GSdx: small optimizations and tried to fix that dx9 fullscreen alt+tab crash

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@1463 96395faa-99c1-11dd-bbfe-3dabce05a288

plugins/GSdx/GPURenderer.cpp

@@ -113,7 +113,17 @@ void GPURenderer::VSync()
 
 	Flush();
 
-	if(!Merge()) return;
+	if(!m_dev->IsLost(true))
+	{
+		if(!Merge())
+		{
+			return;
+		}
+	}
+	else
+	{
+		ResetDevice();
+	}
 
 	// osd 
 
@@ -148,11 +158,6 @@ void GPURenderer::VSync()
 		SetWindowText(m_hWnd, s.c_str());
 	}
 
-	if(m_dev->IsLost())
-	{
-		ResetDevice();
-	}
-
 	GSVector4i r;
 	
 	GetClientRect(m_hWnd, r);

plugins/GSdx/GPURendererSW.cpp

@@ -50,35 +50,15 @@ GSTexture* GPURendererSW::GetOutput()
 	r.right <<= m_scale.x;
 	r.bottom <<= m_scale.y;
 
+	if(m_dev->ResizeTexture(&m_texture, r.width(), r.height()))
+	{
 		// TODO
 		static uint32* buff = (uint32*)_aligned_malloc(m_mem.GetWidth() * m_mem.GetHeight() * sizeof(uint32), 16);
 
 		m_mem.ReadFrame32(r, buff, !!m_env.STATUS.ISRGB24);
 
-	int w = r.width();
-	int h = r.height();
-
-	if(m_texture)
-	{
-		if(m_texture->GetWidth() != w || m_texture->GetHeight() != h)
-		{
-			delete m_texture;
-
-			m_texture = NULL;
-		}
-	}
-
-	if(!m_texture)
-	{
-		m_texture = m_dev->CreateTexture(w, h);
-
-		if(!m_texture)
-		{
-			return NULL;
-		}
-	}
-
 		m_texture->Update(r.rsize(), buff, m_mem.GetWidth() * sizeof(uint32));
+	}
 
 	return m_texture;
 }
@@ -164,8 +144,8 @@ void GPURendererSW::Draw()
 		{
 			GSVector4 p = m_vertices[i].p;
 
-			tl = tl.minv(p);
-			br = br.maxv(p);
+			tl = tl.min(p);
+			br = br.max(p);
 		}
 
 		GSVector4i r = GSVector4i(tl.xyxy(br)).rintersect(data.scissor);

plugins/GSdx/GS.h

@@ -801,7 +801,9 @@ union
 		uint64 _PAD4:30;
 	};
 };
-REG_END
+REG_END2
+	__forceinline bool IsRepeating() {return ((uint32)1 << TW) > (TBW << 6);}
+REG_END2
 
 REG64_(GIFReg, TEX1)
 	uint32 LCM:1;

plugins/GSdx/GSCrc.cpp

@@ -89,6 +89,7 @@ CRC::Game CRC::m_games[] =
 	{0x7D8F539A, SoTC, EU, 0},
 	{0x3122B508, OnePieceGrandAdventure, US, 0},
 	{0x8DF14A24, OnePieceGrandAdventure, Unknown, 0},
+	{0x5D02CC5B, OnePieceGrandBattle, Unknown, 0},
 	{0x6F8545DB, ICO, US, 0},
 	{0xB01A4C95, ICO, JP, 0},
 	{0x5C991F4E, ICO, Unknown, 0},
@@ -143,7 +144,9 @@ CRC::Game CRC::m_games[] =
 	{0x23A97857, StarOcean3, JPUNDUB, 0},
 	{0xCC96CE93, ValkyrieProfile2, US, 0},
 	{0x774DE8E2, ValkyrieProfile2, JP, 0},
+	{0x04CCB600, ValkyrieProfile2, EU, 0},
 	{0x47B9B2FD, RadiataStories, US, 0},
+	{0xE8FCF8EC, SMTNocturne, US, ZWriteMustNotClear}, // saves/reloads z buffer around shadow drawing
 };
 
 hash_map<uint32, CRC::Game*> CRC::m_map;

plugins/GSdx/GSCrc.h

@@ -48,6 +48,7 @@ public:
 		BullyCC,
 		SoTC,
 		OnePieceGrandAdventure,
+		OnePieceGrandBattle,
 		ICO,
 		GT4,
 		WildArms5,
@@ -74,6 +75,7 @@ public:
 		StarOcean3,
 		ValkyrieProfile2,
 		RadiataStories,
+		SMTNocturne,
 		TitleCount,
 	};
 
@@ -95,6 +97,7 @@ public:
 	enum Flags
 	{
 		PointListPalette = 1,
+		ZWriteMustNotClear = 2,
 	};
 
 	struct Game 

plugins/GSdx/GSDevice.cpp

@@ -44,7 +44,7 @@ bool GSDevice::Create(GSWnd* wnd, bool vsync)
 	return true;
 }
 
-bool GSDevice::Reset(int w, int h, bool fs)
+bool GSDevice::Reset(int w, int h, int mode)
 {
 	for(list<GSTexture*>::iterator i = m_pool.begin(); i != m_pool.end(); i++)
 	{
@@ -74,13 +74,15 @@ void GSDevice::Present(const GSVector4i& r, int shader)
 {
 	GSVector4i cr = m_wnd->GetClientRect();
 
-	// Skip Presentation if the surface is invisible (minimized or hidden); prevents DX null swapchain crashes.
-	if(cr.width() == 0 || cr.height() == 0)
-		return;
+	int w = std::max(cr.width(), 1);
+	int h = std::max(cr.height(), 1);
 
-	if(m_backbuffer->GetWidth() != cr.width() || m_backbuffer->GetHeight() != cr.height())
+	if(!m_backbuffer || m_backbuffer->GetWidth() != w || m_backbuffer->GetHeight() != h)
 	{
-		Reset(cr.width(), cr.height(), false);
+		if(!Reset(w, h, DontCare))
+		{
+			return;
+		}
 	}
 
 	ClearRenderTarget(m_backbuffer, 0);
@@ -226,3 +228,21 @@ void GSDevice::Interlace(const GSVector2i& ds, int field, int mode, float yoffse
 		m_current = m_merge;
 	}
 }
+
+bool GSDevice::ResizeTexture(GSTexture** t, int w, int h)
+{
+	if(t == NULL) {ASSERT(0); return false;}
+
+	GSTexture* t2 = *t;
+
+	if(t2 == NULL || t2->GetWidth() != w || t2->GetHeight() != h)
+	{
+		delete t2;
+
+		t2 = CreateTexture(w, h);
+
+		*t = t2;
+	}
+
+	return t2 != NULL;
+}

plugins/GSdx/GSDevice.h

@@ -73,10 +73,11 @@ public:
 
 	void Recycle(GSTexture* t);
 
-	virtual bool Create(GSWnd* wnd, bool vsync);
-	virtual bool Reset(int w, int h, bool fs);
+	enum {Windowed, Fullscreen, DontCare};
 
-	virtual bool IsLost() {return false;}
+	virtual bool Create(GSWnd* wnd, bool vsync);
+	virtual bool Reset(int w, int h, int mode);
+	virtual bool IsLost(bool update = false) {return false;}
 	virtual void Present(const GSVector4i& r, int shader);
 	virtual void Flip() {}
 
@@ -105,6 +106,8 @@ public:
 	void Merge(GSTexture* st[2], GSVector4* sr, GSVector4* dr, const GSVector2i& fs, bool slbg, bool mmod, const GSVector4& c);
 	void Interlace(const GSVector2i& ds, int field, int mode, float yoffset);
 
+	bool ResizeTexture(GSTexture** t, int w, int h);
+
 	template<class T> void PrepareShaderMacro(vector<T>& dst, const T* src, const char* model)
 	{
 		dst.clear();

plugins/GSdx/GSDevice10.cpp

@@ -207,16 +207,16 @@ bool GSDevice10::Create(GSWnd* wnd, bool vsync)
 
 	//
 
-	Reset(1, 1, false);
+	Reset(1, 1, Windowed);
 
 	//
 
 	return true;
 }
 
-bool GSDevice10::Reset(int w, int h, bool fs)
+bool GSDevice10::Reset(int w, int h, int mode)
 {
-	if(!__super::Reset(w, h, fs))
+	if(!__super::Reset(w, h, mode))
 		return false;
 
 	DXGI_SWAP_CHAIN_DESC scd;
@@ -504,7 +504,7 @@ void GSDevice10::IASetVertexBuffer(const void* vertices, size_t stride, size_t c
 		m_vertices.vb = NULL;
 		m_vertices.start = 0;
 		m_vertices.count = 0;
-		m_vertices.limit = max(count * 3 / 2, 10000);
+		m_vertices.limit = std::max<int>(count * 3 / 2, 10000);
 	}
 
 	if(m_vertices.vb == NULL)

plugins/GSdx/GSDevice10.h

@@ -96,8 +96,7 @@ public:
 	virtual ~GSDevice10();
 
 	bool Create(GSWnd* wnd, bool vsync);
-	bool Reset(int w, int h, bool fs);
-
+	bool Reset(int w, int h, int mode);
 	void Flip();
 
 	void BeginScene();

plugins/GSdx/GSDevice11.cpp

@@ -244,16 +244,16 @@ bool GSDevice11::Create(GSWnd* wnd, bool vsync)
 
 	//
 
-	Reset(1, 1, false);
+	Reset(1, 1, Windowed);
 
 	//
 
 	return true;
 }
 
-bool GSDevice11::Reset(int w, int h, bool fs)
+bool GSDevice11::Reset(int w, int h, int mode)
 {
-	if(!__super::Reset(w, h, fs))
+	if(!__super::Reset(w, h, mode))
 		return false;
 
 	DXGI_SWAP_CHAIN_DESC scd;
@@ -541,7 +541,7 @@ void GSDevice11::IASetVertexBuffer(const void* vertices, size_t stride, size_t c
 		m_vertices.vb = NULL;
 		m_vertices.start = 0;
 		m_vertices.count = 0;
-		m_vertices.limit = max(count * 3 / 2, 10000);
+		m_vertices.limit = std::max<int>(count * 3 / 2, 10000);
 	}
 
 	if(m_vertices.vb == NULL)

plugins/GSdx/GSDevice11.h

@@ -99,8 +99,7 @@ public:
 	virtual ~GSDevice11();
 
 	bool Create(GSWnd* wnd, bool vsync);
-	bool Reset(int w, int h, bool fs);
-
+	bool Reset(int w, int h, int mode);
 	void Flip();
 
 	void BeginScene();

plugins/GSdx/GSDevice7.cpp

@@ -25,6 +25,7 @@
 #include "GSDevice7.h"
 
 GSDevice7::GSDevice7()
+	: m_lost(false)
 {
 }
 
@@ -75,14 +76,14 @@ bool GSDevice7::Create(GSWnd* wnd, bool vsync)
 		return false;
 	}
 
-	Reset(1, 1, false);
+	Reset(1, 1, Windowed);
 
 	return true;
 }
 
-bool GSDevice7::Reset(int w, int h, bool fs)
+bool GSDevice7::Reset(int w, int h, int mode)
 {
-	if(!__super::Reset(w, h, fs))
+	if(!__super::Reset(w, h, mode))
 		return false;
 
     DDSURFACEDESC2 desc;
@@ -130,6 +131,8 @@ bool GSDevice7::Reset(int w, int h, bool fs)
 		}
 	}
 
+	m_lost = false;
+
 	return true;
 }
 
@@ -139,9 +142,15 @@ void GSDevice7::Present(const GSVector4i& r, int shader)
 
 	GSVector4i cr = m_wnd->GetClientRect();
 
-	if(m_backbuffer->GetWidth() != cr.width() || m_backbuffer->GetHeight() != cr.height())
+	int w = std::max(cr.width(), 1);
+	int h = std::max(cr.height(), 1);
+
+	if(!m_backbuffer || m_backbuffer->GetWidth() != w || m_backbuffer->GetHeight() != h)
 	{
-		Reset(cr.width(), cr.height(), false);
+		if(!Reset(w, h, DontCare))
+		{
+			return;
+		}
 	}
 
 	CComPtr<IDirectDrawSurface7> backbuffer = *(GSTexture7*)m_backbuffer;
@@ -157,7 +166,10 @@ void GSDevice7::Present(const GSVector4i& r, int shader)
 
 	GSVector4i r2 = r;
 
-	hr = backbuffer->Blt(r2, *(GSTexture7*)m_merge, NULL, DDBLT_WAIT, NULL);
+	if(m_current)
+	{
+		hr = backbuffer->Blt(r2, *(GSTexture7*)m_current, NULL, DDBLT_WAIT, NULL);
+	}
 
 	// if ClearRenderTarget was implemented the parent class could handle these tasks until this point
 
@@ -174,6 +186,8 @@ void GSDevice7::Present(const GSVector4i& r, int shader)
 
 	if(hr == DDERR_SURFACELOST)
 	{
+		m_lost = true;
+
 		// TODO
 
 		HRESULT hr = m_dd->TestCooperativeLevel();

plugins/GSdx/GSDevice7.h

@@ -29,6 +29,7 @@ class GSDevice7 : public GSDevice
 private:
 	CComPtr<IDirectDraw7> m_dd;
 	CComPtr<IDirectDrawSurface7> m_primary;
+	bool m_lost;
 
 	GSTexture* Create(int type, int w, int h, int format);
 
@@ -40,7 +41,7 @@ public:
 	virtual ~GSDevice7();
 
 	bool Create(GSWnd* wnd, bool vsync);
-	bool Reset(int w, int h, bool fs);
-
+	bool Reset(int w, int h, int mode);
+	bool IsLost(bool update) {return m_lost;}
 	void Present(const GSVector4i& r, int shader);
 };

plugins/GSdx/GSDevice9.cpp

@@ -42,6 +42,7 @@ GSDevice9::GSDevice9()
 	, m_bf(0xffffffff)
 	, m_rtv(NULL)
 	, m_dsv(NULL)
+	, m_lost(false)
 {
 	memset(&m_pp, 0, sizeof(m_pp));
 	memset(&m_ddcaps, 0, sizeof(m_ddcaps));
@@ -105,9 +106,7 @@ bool GSDevice9::Create(GSWnd* wnd, bool vsync)
 
 	m_d3d->GetDeviceCaps(D3DADAPTER_DEFAULT, D3DDEVTYPE_HAL, &m_d3dcaps);
 
-	bool fs = theApp.GetConfig("ModeWidth", 0) > 0;
-
-	if(!Reset(1, 1, fs)) return false;
+	if(!Reset(1, 1, theApp.GetConfig("ModeWidth", 0) > 0 ? Fullscreen : Windowed)) return false;
 
 	m_dev->Clear(0, NULL, D3DCLEAR_TARGET, 0, 1.0f, 0);
 
@@ -188,16 +187,23 @@ bool GSDevice9::Create(GSWnd* wnd, bool vsync)
 	return true;
 }
 
-bool GSDevice9::Reset(int w, int h, bool fs)
+bool GSDevice9::Reset(int w, int h, int mode)
 {
-	if(!__super::Reset(w, h, fs))
+	if(!__super::Reset(w, h, mode))
 		return false;
 
 	HRESULT hr;
 
 	if(!m_d3d) return false;
 
-	if(m_swapchain && !fs && m_pp.Windowed)
+	if(mode == DontCare)
+	{
+		mode = m_pp.Windowed ? Windowed : Fullscreen;
+	}
+
+	if(!m_lost)
+	{
+		if(m_swapchain && mode != Fullscreen && m_pp.Windowed)
 		{
 			m_swapchain = NULL;
 
@@ -214,9 +220,15 @@ bool GSDevice9::Reset(int w, int h, bool fs)
 
 			return true;
 		}
+	}
 
 	m_swapchain = NULL;
 
+	m_vertices.vb = NULL;
+	m_vertices.vb_old = NULL;
+	m_vertices.start = 0;
+	m_vertices.count = 0;
+
 	if(m_vs_cb) _aligned_free(m_vs_cb);
 	if(m_ps_cb) _aligned_free(m_ps_cb);
 
@@ -258,7 +270,7 @@ bool GSDevice9::Reset(int w, int h, bool fs)
 	int mh = theApp.GetConfig("ModeHeight", 0);
 	int mrr = theApp.GetConfig("ModeRefreshRate", 0);
 
-	if(fs && mw > 0 && mh > 0 && mrr >= 0)
+	if(mode == Fullscreen && mw > 0 && mh > 0 && mrr >= 0)
 	{
 		m_pp.Windowed = FALSE;
 		m_pp.BackBufferWidth = mw;
@@ -324,27 +336,39 @@ bool GSDevice9::Reset(int w, int h, bool fs)
 	return true;
 }
 
-bool GSDevice9::IsLost()
+bool GSDevice9::IsLost(bool update)
+{
+	if(!m_lost || update)
 	{
 		HRESULT hr = m_dev->TestCooperativeLevel();
 
-	return hr == D3DERR_DEVICELOST || hr == D3DERR_DEVICENOTRESET;
+		m_lost = hr == D3DERR_DEVICELOST || hr == D3DERR_DEVICENOTRESET;
+	}
+
+	return m_lost;
 }
 
 void GSDevice9::Flip()
 {
 	m_dev->EndScene();
 
+	HRESULT hr;
+
 	if(m_swapchain)
 	{
-		m_swapchain->Present(NULL, NULL, NULL, NULL, 0);
+		hr = m_swapchain->Present(NULL, NULL, NULL, NULL, 0);
 	}
 	else
 	{
-		m_dev->Present(NULL, NULL, NULL, NULL);
+		hr = m_dev->Present(NULL, NULL, NULL, NULL);
 	}
 
 	m_dev->BeginScene();
+
+	if(FAILED(hr))
+	{
+		m_lost = true;
+	}
 }
 
 void GSDevice9::BeginScene()
@@ -663,7 +687,7 @@ void GSDevice9::IASetVertexBuffer(const void* vertices, size_t stride, size_t co
 		m_vertices.vb = NULL;
 		m_vertices.start = 0;
 		m_vertices.count = 0;
-		m_vertices.limit = max(count * 3 / 2, 10000);
+		m_vertices.limit = std::max<int>(count * 3 / 2, 10000);
 	}
 
 	if(m_vertices.vb == NULL)

plugins/GSdx/GSDevice9.h

@@ -95,6 +95,7 @@ private:
 	CComPtr<IDirect3D9> m_d3d;
 	CComPtr<IDirect3DDevice9> m_dev;
 	CComPtr<IDirect3DSwapChain9> m_swapchain;
+	bool m_lost;
 
 	struct
 	{
@@ -132,9 +133,8 @@ public:
 	virtual ~GSDevice9();
 
 	bool Create(GSWnd* wnd, bool vsync);
-	bool Reset(int w, int h, bool fs);
-
-	bool IsLost();
+	bool Reset(int w, int h, int mode);
+	bool IsLost(bool update);
 	void Flip();
 
 	void BeginScene();

plugins/GSdx/GSDeviceNull.cpp

@@ -29,14 +29,14 @@ bool GSDeviceNull::Create(GSWnd* wnd, bool vsync)
 		return false;
 	}
 
-	Reset(1, 1, false);
+	Reset(1, 1, Windowed);
 
 	return true;
 }
 
-bool GSDeviceNull::Reset(int w, int h, bool fs)
+bool GSDeviceNull::Reset(int w, int h, int mode)
 {
-	if(!__super::Reset(w, h, fs))
+	if(!__super::Reset(w, h, mode))
 		return false;
 
 	return true;

plugins/GSdx/GSDeviceNull.h

@@ -36,5 +36,5 @@ public:
 	GSDeviceNull() {}
 
 	bool Create(GSWnd* wnd, bool vsync);
-	bool Reset(int w, int h, bool fs);
+	bool Reset(int w, int h, int mode);
 };

plugins/GSdx/GSDeviceOGL.cpp

@@ -150,14 +150,14 @@ bool GSDeviceOGL::Create(GSWnd* wnd, bool vsync)
 */
 	GSVector4i r = wnd->GetClientRect();
 
-	Reset(r.width(), r.height(), false);
+	Reset(r.width(), r.height(), Windowed);
 
 	return true;
 }
 
-bool GSDeviceOGL::Reset(int w, int h, bool fs)
+bool GSDeviceOGL::Reset(int w, int h, int mode)
 {
-	if(!__super::Reset(w, h, fs))
+	if(!__super::Reset(w, h, mode))
 		return false;
 
 	glCullFace(GL_FRONT_AND_BACK); CheckError();
@@ -361,7 +361,7 @@ void GSDeviceOGL::IASetVertexBuffer(const void* vertices, size_t stride, size_t
 	{
 		m_vertices.start = 0;
 		m_vertices.count = 0;
-		m_vertices.limit = max(count * 3 / 2, 10000);
+		m_vertices.limit = std::max<int>(count * 3 / 2, 10000);
 
 		growbuffer = true;
 	}

plugins/GSdx/GSDeviceOGL.h

@@ -110,8 +110,7 @@ public:
 	virtual ~GSDeviceOGL();
 
 	bool Create(GSWnd* wnd, bool vsync);
-	bool Reset(int w, int h, bool fs);
-
+	bool Reset(int w, int h, int mode);
 	void Present(const GSVector4i& r, int shader);
 	void Flip();
 

plugins/GSdx/GSDrawScanline.cpp

@@ -48,8 +48,8 @@ void GSDrawScanline::BeginDraw(const GSRasterizerData* data, Functions* f)
 	m_env.vm = p->vm;
 	m_env.fbr = p->fbo->row;
 	m_env.zbr = p->zbo->row;
-	m_env.fbc = p->fbo->col;
-	m_env.zbc = p->zbo->col;
+	m_env.fbc = p->fbo->col[0];
+	m_env.zbc = p->zbo->col[0];
 	m_env.fzbr = p->fzbo->row;
 	m_env.fzbc = p->fzbo->col;
 	m_env.fm = GSVector4i(p->fm);
@@ -116,8 +116,8 @@ void GSDrawScanline::BeginDraw(const GSRasterizerData* data, Functions* f)
 			m_env.t.mask.u32[0] = 0; 
 			break;
 		case CLAMP_REGION_CLAMP: 
-			m_env.t.min.u16[0] = min(context->CLAMP.MINU, tw - 1);
-			m_env.t.max.u16[0] = min(context->CLAMP.MAXU, tw - 1);
+			m_env.t.min.u16[0] = std::min<int>(context->CLAMP.MINU, tw - 1);
+			m_env.t.max.u16[0] = std::min<int>(context->CLAMP.MAXU, tw - 1);
 			m_env.t.mask.u32[0] = 0; 
 			break;
 		case CLAMP_REGION_REPEAT: 
@@ -142,8 +142,8 @@ void GSDrawScanline::BeginDraw(const GSRasterizerData* data, Functions* f)
 			m_env.t.mask.u32[2] = 0; 
 			break;
 		case CLAMP_REGION_CLAMP: 
-			m_env.t.min.u16[4] = min(context->CLAMP.MINV, th - 1);
-			m_env.t.max.u16[4] = min(context->CLAMP.MAXV, th - 1); // ffx anima summon scene, when the anchor appears (th = 256, maxv > 256)
+			m_env.t.min.u16[4] = std::min<int>(context->CLAMP.MINV, th - 1);
+			m_env.t.max.u16[4] = std::min<int>(context->CLAMP.MAXV, th - 1); // ffx anima summon scene, when the anchor appears (th = 256, maxv > 256)
 			m_env.t.mask.u32[2] = 0; 
 			break;
 		case CLAMP_REGION_REPEAT: 
@@ -224,22 +224,22 @@ void GSDrawScanline::DrawSolidRect(const GSVector4i& r, const GSVertexSW& v)
 		{
 			if(m == 0)
 			{
-				DrawSolidRectT<uint32, false>(m_env.zbr, m_env.zbc[0], r, z, m);
+				DrawSolidRectT<uint32, false>(m_env.zbr, m_env.zbc, r, z, m);
 			}
 			else
 			{
-				DrawSolidRectT<uint32, true>(m_env.zbr, m_env.zbc[0], r, z, m);
+				DrawSolidRectT<uint32, true>(m_env.zbr, m_env.zbc, r, z, m);
 			}
 		}
 		else
 		{
 			if(m == 0)
 			{
-				DrawSolidRectT<uint16, false>(m_env.zbr, m_env.zbc[0], r, z, m);
+				DrawSolidRectT<uint16, false>(m_env.zbr, m_env.zbc, r, z, m);
 			}
 			else
 			{
-				DrawSolidRectT<uint16, true>(m_env.zbr, m_env.zbc[0], r, z, m);
+				DrawSolidRectT<uint16, true>(m_env.zbr, m_env.zbc, r, z, m);
 			}
 		}
 	}
@@ -259,11 +259,11 @@ void GSDrawScanline::DrawSolidRect(const GSVector4i& r, const GSVertexSW& v)
 		{
 			if(m == 0)
 			{
-				DrawSolidRectT<uint32, false>(m_env.fbr, m_env.fbc[0], r, c, m);
+				DrawSolidRectT<uint32, false>(m_env.fbr, m_env.fbc, r, c, m);
 			}
 			else
 			{
-				DrawSolidRectT<uint32, true>(m_env.fbr, m_env.fbc[0], r, c, m);
+				DrawSolidRectT<uint32, true>(m_env.fbr, m_env.fbc, r, c, m);
 			}
 		}
 		else
@@ -272,18 +272,18 @@ void GSDrawScanline::DrawSolidRect(const GSVector4i& r, const GSVertexSW& v)
 
 			if(m == 0)
 			{
-				DrawSolidRectT<uint16, false>(m_env.fbr, m_env.fbc[0], r, c, m);
+				DrawSolidRectT<uint16, false>(m_env.fbr, m_env.fbc, r, c, m);
 			}
 			else
 			{
-				DrawSolidRectT<uint16, true>(m_env.fbr, m_env.fbc[0], r, c, m);
+				DrawSolidRectT<uint16, true>(m_env.fbr, m_env.fbc, r, c, m);
 			}
 		}
 	}
 }
 
 template<class T, bool masked> 
-void GSDrawScanline::DrawSolidRectT(const GSVector4i* row, int* col, const GSVector4i& r, uint32 c, uint32 m)
+void GSDrawScanline::DrawSolidRectT(const int* row, int* col, const GSVector4i& r, uint32 c, uint32 m)
 {
 	if(m == 0xffffffff) return;
 
@@ -320,13 +320,13 @@ void GSDrawScanline::DrawSolidRectT(const GSVector4i* row, int* col, const GSVec
 }
 
 template<class T, bool masked> 
-void GSDrawScanline::FillRect(const GSVector4i* row, int* col, const GSVector4i& r, uint32 c, uint32 m)
+void GSDrawScanline::FillRect(const int* row, int* col, const GSVector4i& r, uint32 c, uint32 m)
 {
 	if(r.x >= r.z) return;
 
 	for(int y = r.y; y < r.w; y++)
 	{
-		uint32 base = row[y].x;
+		uint32 base = row[y];
 
 		for(int x = r.x; x < r.z; x++)
 		{
@@ -338,13 +338,13 @@ void GSDrawScanline::FillRect(const GSVector4i* row, int* col, const GSVector4i&
 }
 
 template<class T, bool masked> 
-void GSDrawScanline::FillBlock(const GSVector4i* row, int* col, const GSVector4i& r, const GSVector4i& c, const GSVector4i& m)
+void GSDrawScanline::FillBlock(const int* row, int* col, const GSVector4i& r, const GSVector4i& c, const GSVector4i& m)
 {
 	if(r.x >= r.z) return;
 
 	for(int y = r.y; y < r.w; y += 8)
 	{
-		uint32 base = row[y].x;
+		uint32 base = row[y];
 
 		for(int x = r.x; x < r.z; x += 8 * 4 / sizeof(T))
 		{

plugins/GSdx/GSDrawScanline.h

@@ -60,13 +60,13 @@ class GSDrawScanline : public GSAlignedClass<16>, public IDrawScanline
 	void DrawSolidRect(const GSVector4i& r, const GSVertexSW& v);
 
 	template<class T, bool masked> 
-	void DrawSolidRectT(const GSVector4i* row, int* col, const GSVector4i& r, uint32 c, uint32 m);
+	void DrawSolidRectT(const int* row, int* col, const GSVector4i& r, uint32 c, uint32 m);
 
 	template<class T, bool masked> 
-	__forceinline void FillRect(const GSVector4i* row, int* col, const GSVector4i& r, uint32 c, uint32 m);
+	__forceinline void FillRect(const int* row, int* col, const GSVector4i& r, uint32 c, uint32 m);
 
 	template<class T, bool masked> 
-	__forceinline void FillBlock(const GSVector4i* row, int* col, const GSVector4i& r, const GSVector4i& c, const GSVector4i& m);
+	__forceinline void FillBlock(const int* row, int* col, const GSVector4i& r, const GSVector4i& c, const GSVector4i& m);
 
 protected:
 	GSState* m_state;

plugins/GSdx/GSLocalMemory.cpp

@@ -30,17 +30,19 @@
 #define ASSERT_BLOCK(r, w, h) \
 	ASSERT((r).width() >= w && (r).height() >= h && !((r).left & (w - 1)) && !((r).top & (h - 1)) && !((r).right & (w - 1)) && !((r).bottom & (h - 1))); \
 
-#define FOREACH_BLOCK_START(w, h, bpp, format) \
-	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[format]; \
-	uint32 bp = TEX0.TBP0; \
-	uint32 bw = TEX0.TBW; \
-	int offset = dstpitch * h - r.width() * bpp / 8; \
-	for(int y = r.top, ye = r.bottom; y < ye; y += h, dst += offset) \
-	{ ASSERT_BLOCK(r, w, h); \
-		uint32 base = psm.bn(0, y, bp, bw); \
-		for(int x = r.left, xe = r.right; x < xe; x += w, dst += w * bpp / 8) \
+#define FOREACH_BLOCK_START(r, w, h, bpp, psm) \
+	ASSERT_BLOCK(r, w, h); \
+	const GSLocalMemory::BlockOffset* RESTRICT _bo = GetBlockOffset(TEX0.TBP0, TEX0.TBW, psm); \
+	GSVector4i _r = r >> 3; \
+	uint8* _dst = dst - _r.left * bpp; \
+	int _offset = dstpitch * h; \
+	for(int y = _r.top; y < _r.bottom; y += h >> 3, _dst += _offset) \
 	{ \
-			const uint8* src = BlockPtr(base + psm.blockOffset[x >> 3]); \
+		uint32 _base = _bo->row[y]; \
+		for(int x = _r.left; x < _r.right; x += w >> 3) \
+		{ \
+			const uint8* src = BlockPtr(_base + _bo->col[x]); \
+			uint8* dst = &_dst[x * bpp]; \
 
 #define FOREACH_BLOCK_END }}
 
@@ -463,37 +465,65 @@ GSLocalMemory::~GSLocalMemory()
 {
 	VirtualFree(m_vm8, 0, MEM_RELEASE);
 
-	for(hash_map<uint32, Offset*>::iterator i = m_omap.begin(); i != m_omap.end(); i++)
+	for(hash_map<uint32, BlockOffset*>::iterator i = m_bomap.begin(); i != m_bomap.end(); i++)
 	{
-		Offset* o = i->second;
-		
-		_aligned_free(o->col[0]);
-
-		_aligned_free(o);
+		_aligned_free(i->second);
 	}
 
-	for(hash_map<uint32, Offset4*>::iterator i = m_o4map.begin(); i != m_o4map.end(); i++)
+	for(hash_map<uint32, PixelOffset*>::iterator i = m_pomap.begin(); i != m_pomap.end(); i++)
+	{
+		_aligned_free(i->second);
+	}
+
+	for(hash_map<uint32, PixelOffset4*>::iterator i = m_po4map.begin(); i != m_po4map.end(); i++)
 	{
 		_aligned_free(i->second);
 	}
 }
 
-GSLocalMemory::Offset* GSLocalMemory::GetOffset(uint32 bp, uint32 bw, uint32 psm)
+GSLocalMemory::BlockOffset* GSLocalMemory::GetBlockOffset(uint32 bp, uint32 bw, uint32 psm)
 {
-	if(bw == 0) {ASSERT(0); return NULL;}
-
-	ASSERT(m_psm[psm].bpp > 8); // only for 16/24/32/8h/4hh/4hl formats where all columns are the same
-
 	uint32 hash = bp | (bw << 14) | (psm << 20);
 
-	hash_map<uint32, Offset*>::iterator i = m_omap.find(hash);
+	hash_map<uint32, BlockOffset*>::iterator i = m_bomap.find(hash);
 
-	if(i != m_omap.end())
+	if(i != m_bomap.end())
 	{
 		return i->second;
 	}
 
-	Offset* o = (Offset*)_aligned_malloc(sizeof(Offset), 16);
+	BlockOffset* o = (BlockOffset*)_aligned_malloc(sizeof(BlockOffset), 16);
+
+	o->hash = hash;
+
+	pixelAddress bn = m_psm[psm].bn;
+
+	for(int i = 0; i < 256; i++)
+	{
+		o->row[i] = (int)bn(0, i << 3, bp, bw);
+	}
+
+	o->col = m_psm[psm].blockOffset;
+
+	m_bomap[hash] = o;
+
+	return o;
+}
+
+GSLocalMemory::PixelOffset* GSLocalMemory::GetPixelOffset(uint32 bp, uint32 bw, uint32 psm)
+{
+	if(bw == 0) {ASSERT(0); return NULL;}
+
+	uint32 hash = bp | (bw << 14) | (psm << 20);
+
+	hash_map<uint32, PixelOffset*>::iterator i = m_pomap.find(hash);
+
+	if(i != m_pomap.end())
+	{
+		return i->second;
+	}
+
+	PixelOffset* o = (PixelOffset*)_aligned_malloc(sizeof(PixelOffset), 16);
 
 	o->hash = hash;
 
@@ -501,24 +531,20 @@ GSLocalMemory::Offset* GSLocalMemory::GetOffset(uint32 bp, uint32 bw, uint32 psm
 
 	for(int i = 0; i < 2048; i++)
 	{
-		o->row[i] = GSVector4i((int)pa(0, i, bp, bw));
+		o->row[i] = (int)pa(0, i, bp, bw);
 	}
 
-	int* p = (int*)_aligned_malloc(sizeof(int) * (2048 + 3) * 4, 16);
-
-	for(int i = 0; i < 4; i++)
+	for(int i = 0; i < 8; i++)
 	{
-		o->col[i] = &p[2048 * i + ((4 - (i & 3)) & 3)];
-
-		memcpy(o->col[i], m_psm[psm].rowOffset[0], sizeof(int) * 2048);
+		o->col[i] = m_psm[psm].rowOffset[i];
 	}
 
-	m_omap[hash] = o;
+	m_pomap[hash] = o;
 
 	return o;
 }
 
-GSLocalMemory::Offset4* GSLocalMemory::GetOffset4(const GIFRegFRAME& FRAME, const GIFRegZBUF& ZBUF)
+GSLocalMemory::PixelOffset4* GSLocalMemory::GetPixelOffset4(const GIFRegFRAME& FRAME, const GIFRegZBUF& ZBUF)
 {
 	uint32 fbp = FRAME.Block();
 	uint32 zbp = ZBUF.Block();
@@ -535,14 +561,14 @@ GSLocalMemory::Offset4* GSLocalMemory::GetOffset4(const GIFRegFRAME& FRAME, cons
 
 	uint32 hash = (FRAME.FBP << 0) | (ZBUF.ZBP << 9) | (bw << 18) | (fpsm_hash << 24) | (zpsm_hash << 28);
 
-	hash_map<uint32, Offset4*>::iterator i = m_o4map.find(hash);
+	hash_map<uint32, PixelOffset4*>::iterator i = m_po4map.find(hash);
 
-	if(i != m_o4map.end())
+	if(i != m_po4map.end())
 	{
 		return i->second;
 	}
 
-	Offset4* o = (Offset4*)_aligned_malloc(sizeof(Offset4), 16);
+	PixelOffset4* o = (PixelOffset4*)_aligned_malloc(sizeof(PixelOffset4), 16);
 
 	o->hash = hash;
 
@@ -564,7 +590,7 @@ GSLocalMemory::Offset4* GSLocalMemory::GetOffset4(const GIFRegFRAME& FRAME, cons
 		o->col[i].y = m_psm[zpsm].rowOffset[0][i * 4] << zs;
 	}
 
-	m_o4map[hash] = o;
+	m_po4map[hash] = o;
 
 	return o;
 }
@@ -1436,20 +1462,20 @@ void GSLocalMemory::ReadImageX(int& tx, int& ty, uint8* dst, int len, GIFRegBITB
 
 ///////////////////
 
-void GSLocalMemory::ReadTexture32(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture32(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(8, 8, 32, PSM_PSMCT32)
+	FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMCT32)
 	{
 		ReadBlock32<true>(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture24(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture24(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	if(TEXA.AEM)
 	{
-		FOREACH_BLOCK_START(8, 8, 32, PSM_PSMCT24)
+		FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMCT24)
 		{
 			ReadAndExpandBlock24<true>(src, dst, dstpitch, TEXA);
 		}
@@ -1457,7 +1483,7 @@ void GSLocalMemory::ReadTexture24(const GSVector4i& r, uint8* dst, int dstpitch,
 	}
 	else
 	{
-		FOREACH_BLOCK_START(8, 8, 32, PSM_PSMCT24)
+		FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMCT24)
 		{
 			ReadAndExpandBlock24<false>(src, dst, dstpitch, TEXA);
 		}
@@ -1465,11 +1491,11 @@ void GSLocalMemory::ReadTexture24(const GSVector4i& r, uint8* dst, int dstpitch,
 	}
 }
 
-void GSLocalMemory::ReadTexture16(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	__declspec(align(16)) uint16 block[16 * 8];
 
-	FOREACH_BLOCK_START(16, 8, 32, PSM_PSMCT16)
+	FOREACH_BLOCK_START(r, 16, 8, 32, PSM_PSMCT16)
 	{
 		ReadBlock16<true>(src, (uint8*)block, sizeof(block) / 8);
 
@@ -1478,11 +1504,11 @@ void GSLocalMemory::ReadTexture16(const GSVector4i& r, uint8* dst, int dstpitch,
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture16S(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16S(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	__declspec(align(16)) uint16 block[16 * 8];
 
-	FOREACH_BLOCK_START(16, 8, 32, PSM_PSMCT16S)
+	FOREACH_BLOCK_START(r, 16, 8, 32, PSM_PSMCT16S)
 	{
 		ReadBlock16<true>(src, (uint8*)block, sizeof(block) / 8);
 
@@ -1491,75 +1517,75 @@ void GSLocalMemory::ReadTexture16S(const GSVector4i& r, uint8* dst, int dstpitch
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture8(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture8(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
-	FOREACH_BLOCK_START(16, 16, 32, PSM_PSMT8)
+	FOREACH_BLOCK_START(r, 16, 16, 32, PSM_PSMT8)
 	{
 		ReadAndExpandBlock8_32(src, dst, dstpitch, pal);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture4(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint64* pal = m_clut;
 
-	FOREACH_BLOCK_START(32, 16, 32, PSM_PSMT4)
+	FOREACH_BLOCK_START(r, 32, 16, 32, PSM_PSMT4)
 	{
 		ReadAndExpandBlock4_32(src, dst, dstpitch, pal);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture8H(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture8H(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
-	FOREACH_BLOCK_START(8, 8, 32, PSM_PSMT8H)
+	FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMT8H)
 	{
 		ReadAndExpandBlock8H_32(src, dst, dstpitch, pal);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture4HL(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4HL(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
-	FOREACH_BLOCK_START(8, 8, 32, PSM_PSMT4HL)
+	FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMT4HL)
 	{
 		ReadAndExpandBlock4HL_32(src, dst, dstpitch, pal);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture4HH(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4HH(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
-	FOREACH_BLOCK_START(8, 8, 32, PSM_PSMT4HH)
+	FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMT4HH)
 	{
 		ReadAndExpandBlock4HH_32(src, dst, dstpitch, pal);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture32Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture32Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(8, 8, 32, PSM_PSMZ32)
+	FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMZ32)
 	{
 		ReadBlock32<true>(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture24Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture24Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	if(TEXA.AEM)
 	{
-		FOREACH_BLOCK_START(8, 8, 32, PSM_PSMZ24)
+		FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMZ24)
 		{
 			ReadAndExpandBlock24<true>(src, dst, dstpitch, TEXA);
 		}
@@ -1567,7 +1593,7 @@ void GSLocalMemory::ReadTexture24Z(const GSVector4i& r, uint8* dst, int dstpitch
 	}
 	else
 	{
-		FOREACH_BLOCK_START(8, 8, 32, PSM_PSMZ24)
+		FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMZ24)
 		{
 			ReadAndExpandBlock24<false>(src, dst, dstpitch, TEXA);
 		}
@@ -1575,11 +1601,11 @@ void GSLocalMemory::ReadTexture24Z(const GSVector4i& r, uint8* dst, int dstpitch
 	}
 }
 
-void GSLocalMemory::ReadTexture16Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	__declspec(align(16)) uint16 block[16 * 8];
 
-	FOREACH_BLOCK_START(16, 8, 32, PSM_PSMZ16)
+	FOREACH_BLOCK_START(r, 16, 8, 32, PSM_PSMZ16)
 	{
 		ReadBlock16<true>(src, (uint8*)block, sizeof(block) / 8);
 
@@ -1588,11 +1614,11 @@ void GSLocalMemory::ReadTexture16Z(const GSVector4i& r, uint8* dst, int dstpitch
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture16SZ(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16SZ(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	__declspec(align(16)) uint16 block[16 * 8];
 
-	FOREACH_BLOCK_START(16, 8, 32, PSM_PSMZ16S)
+	FOREACH_BLOCK_START(r, 16, 8, 32, PSM_PSMZ16S)
 	{
 		ReadBlock16<true>(src, (uint8*)block, sizeof(block) / 8);
 
@@ -1737,31 +1763,31 @@ void GSLocalMemory::ReadTexture(const GSVector4i& r, uint8* dst, int dstpitch, c
 }
 ///////////////////
 
-void GSLocalMemory::ReadTexture16NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(16, 8, 16, PSM_PSMCT16)
+	FOREACH_BLOCK_START(r, 16, 8, 16, PSM_PSMCT16)
 	{
 		ReadBlock16<true>(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture16SNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16SNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(16, 8, 16, PSM_PSMCT16S)
+	FOREACH_BLOCK_START(r, 16, 8, 16, PSM_PSMCT16S)
 	{
 		ReadBlock16<true>(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture8NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture8NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
 	if(TEX0.CPSM == PSM_PSMCT32 || TEX0.CPSM == PSM_PSMCT24)
 	{
-		FOREACH_BLOCK_START(16, 16, 32, PSM_PSMT8)
+		FOREACH_BLOCK_START(r, 16, 16, 32, PSM_PSMT8)
 		{
 			ReadAndExpandBlock8_32(src, dst, dstpitch, pal);
 		}
@@ -1773,7 +1799,7 @@ void GSLocalMemory::ReadTexture8NP(const GSVector4i& r, uint8* dst, int dstpitch
 
 		__declspec(align(16)) uint8 block[16 * 16];
 
-		FOREACH_BLOCK_START(16, 16, 16, PSM_PSMT8)
+		FOREACH_BLOCK_START(r, 16, 16, 16, PSM_PSMT8)
 		{
 			ReadBlock8<true>(src, (uint8*)block, sizeof(block) / 16);
 
@@ -1783,13 +1809,13 @@ void GSLocalMemory::ReadTexture8NP(const GSVector4i& r, uint8* dst, int dstpitch
 	}
 }
 
-void GSLocalMemory::ReadTexture4NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint64* pal = m_clut;
 
 	if(TEX0.CPSM == PSM_PSMCT32 || TEX0.CPSM == PSM_PSMCT24)
 	{
-		FOREACH_BLOCK_START(32, 16, 32, PSM_PSMT4)
+		FOREACH_BLOCK_START(r, 32, 16, 32, PSM_PSMT4)
 		{
 			ReadAndExpandBlock4_32(src, dst, dstpitch, pal);
 		}
@@ -1801,7 +1827,7 @@ void GSLocalMemory::ReadTexture4NP(const GSVector4i& r, uint8* dst, int dstpitch
 
 		__declspec(align(16)) uint8 block[(32 / 2) * 16];
 
-		FOREACH_BLOCK_START(32, 16, 16, PSM_PSMT4)
+		FOREACH_BLOCK_START(r, 32, 16, 16, PSM_PSMT4)
 		{
 			ReadBlock4<true>(src, (uint8*)block, sizeof(block) / 16);
 
@@ -1811,13 +1837,13 @@ void GSLocalMemory::ReadTexture4NP(const GSVector4i& r, uint8* dst, int dstpitch
 	}
 }
 
-void GSLocalMemory::ReadTexture8HNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture8HNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
 	if(TEX0.CPSM == PSM_PSMCT32 || TEX0.CPSM == PSM_PSMCT24)
 	{
-		FOREACH_BLOCK_START(8, 8, 32, PSM_PSMT8H)
+		FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMT8H)
 		{
 			ReadAndExpandBlock8H_32(src, dst, dstpitch, pal);
 		}
@@ -1829,7 +1855,7 @@ void GSLocalMemory::ReadTexture8HNP(const GSVector4i& r, uint8* dst, int dstpitc
 
 		__declspec(align(16)) uint32 block[8 * 8];
 
-		FOREACH_BLOCK_START(8, 8, 16, PSM_PSMT8H)
+		FOREACH_BLOCK_START(r, 8, 8, 16, PSM_PSMT8H)
 		{
 			ReadBlock32<true>(src, (uint8*)block, sizeof(block) / 8);
 
@@ -1839,13 +1865,13 @@ void GSLocalMemory::ReadTexture8HNP(const GSVector4i& r, uint8* dst, int dstpitc
 	}
 }
 
-void GSLocalMemory::ReadTexture4HLNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4HLNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
 	if(TEX0.CPSM == PSM_PSMCT32 || TEX0.CPSM == PSM_PSMCT24)
 	{
-		FOREACH_BLOCK_START(8, 8, 32, PSM_PSMT4HL)
+		FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMT4HL)
 		{
 			ReadAndExpandBlock4HL_32(src, dst, dstpitch, pal);
 		}
@@ -1857,7 +1883,7 @@ void GSLocalMemory::ReadTexture4HLNP(const GSVector4i& r, uint8* dst, int dstpit
 
 		__declspec(align(16)) uint32 block[8 * 8];
 
-		FOREACH_BLOCK_START(8, 8, 16, PSM_PSMT4HL)
+		FOREACH_BLOCK_START(r, 8, 8, 16, PSM_PSMT4HL)
 		{
 			ReadBlock32<true>(src, (uint8*)block, sizeof(block) / 8);
 
@@ -1867,13 +1893,13 @@ void GSLocalMemory::ReadTexture4HLNP(const GSVector4i& r, uint8* dst, int dstpit
 	}
 }
 
-void GSLocalMemory::ReadTexture4HHNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4HHNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
 	const uint32* pal = m_clut;
 
 	if(TEX0.CPSM == PSM_PSMCT32 || TEX0.CPSM == PSM_PSMCT24)
 	{
-		FOREACH_BLOCK_START(8, 8, 32, PSM_PSMT4HH)
+		FOREACH_BLOCK_START(r, 8, 8, 32, PSM_PSMT4HH)
 		{
 			ReadAndExpandBlock4HH_32(src, dst, dstpitch, pal);
 		}
@@ -1885,7 +1911,7 @@ void GSLocalMemory::ReadTexture4HHNP(const GSVector4i& r, uint8* dst, int dstpit
 
 		__declspec(align(16)) uint32 block[8 * 8];
 
-		FOREACH_BLOCK_START(8, 8, 16, PSM_PSMT4HH)
+		FOREACH_BLOCK_START(r, 8, 8, 16, PSM_PSMT4HH)
 		{
 			ReadBlock32<true>(src, (uint8*)block, sizeof(block) / 8);
 
@@ -1895,18 +1921,18 @@ void GSLocalMemory::ReadTexture4HHNP(const GSVector4i& r, uint8* dst, int dstpit
 	}
 }
 
-void GSLocalMemory::ReadTexture16ZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16ZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(16, 8, 16, PSM_PSMZ16)
+	FOREACH_BLOCK_START(r, 16, 8, 16, PSM_PSMZ16)
 	{
 		ReadBlock16<true>(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture16SZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture16SZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(16, 8, 16, PSM_PSMZ16S)
+	FOREACH_BLOCK_START(r, 16, 8, 16, PSM_PSMZ16S)
 	{
 		ReadBlock16<true>(src, dst, dstpitch);
 	}
@@ -1959,45 +1985,45 @@ void GSLocalMemory::ReadTextureNP(const GSVector4i& r, uint8* dst, int dstpitch,
 
 // 32/8
 
-void GSLocalMemory::ReadTexture8P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture8P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(16, 16, 8, PSM_PSMT8)
+	FOREACH_BLOCK_START(r, 16, 16, 8, PSM_PSMT8)
 	{
 		ReadBlock8<true>(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture4P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(32, 16, 8, PSM_PSMT4)
+	FOREACH_BLOCK_START(r, 32, 16, 8, PSM_PSMT4)
 	{
 		ReadBlock4P(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture8HP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture8HP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(8, 8, 8, PSM_PSMT8H)
+	FOREACH_BLOCK_START(r, 8, 8, 8, PSM_PSMT8H)
 	{
 		ReadBlock8HP(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture4HLP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4HLP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(8, 8, 8, PSM_PSMT4HL)
+	FOREACH_BLOCK_START(r, 8, 8, 8, PSM_PSMT4HL)
 	{
 		ReadBlock4HLP(src, dst, dstpitch);
 	}
 	FOREACH_BLOCK_END
 }
 
-void GSLocalMemory::ReadTexture4HHP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const
+void GSLocalMemory::ReadTexture4HHP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA)
 {
-	FOREACH_BLOCK_START(8, 8, 8, PSM_PSMT4HH)
+	FOREACH_BLOCK_START(r, 8, 8, 8, PSM_PSMT4HH)
 	{
 		ReadBlock4HHP(src, dst, dstpitch);
 	}

plugins/GSdx/GSLocalMemory.h

@@ -43,7 +43,7 @@ public:
 	typedef uint32 (GSLocalMemory::*readTexelAddr)(uint32 addr, const GIFRegTEXA& TEXA) const;
 	typedef void (GSLocalMemory::*writeImage)(int& tx, int& ty, uint8* src, int len, GIFRegBITBLTBUF& BITBLTBUF, GIFRegTRXPOS& TRXPOS, GIFRegTRXREG& TRXREG);
 	typedef void (GSLocalMemory::*readImage)(int& tx, int& ty, uint8* dst, int len, GIFRegBITBLTBUF& BITBLTBUF, GIFRegTRXPOS& TRXPOS, GIFRegTRXREG& TRXREG) const;
-	typedef void (GSLocalMemory::*readTexture)(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
+	typedef void (GSLocalMemory::*readTexture)(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
 	typedef void (GSLocalMemory::*readTextureBlock)(uint32 bp, uint8* dst, int dstpitch, const GIFRegTEXA& TEXA) const;
 
 	typedef union 
@@ -81,14 +81,21 @@ public:
 
 	GSClut m_clut;
 
-	struct Offset
+	struct BlockOffset
 	{
-		GSVector4i row[2048]; // 0 | 0 | 0 | 0
-		int* col[4]; // x | x+1 | x+2 | x+3
+		int row[256]; // yn (n = 0 8 16 ...)
+		int* col; // blockOffset*
 		uint32 hash;
 	};
 
-	struct Offset4
+	struct PixelOffset
+	{
+		int row[2048]; // yn (n = 0 1 2 ...)
+		int* col[8]; // rowOffset*
+		uint32 hash;
+	};
+
+	struct PixelOffset4
 	{
 		// 16 bit offsets (m_vm16[...])
 
@@ -141,15 +148,17 @@ protected:
 
 	//
 
-	hash_map<uint32, Offset*> m_omap;
-	hash_map<uint32, Offset4*> m_o4map;
+	hash_map<uint32, BlockOffset*> m_bomap;
+	hash_map<uint32, PixelOffset*> m_pomap;
+	hash_map<uint32, PixelOffset4*> m_po4map;
 
 public:
 	GSLocalMemory();
 	virtual ~GSLocalMemory();
 
-	Offset* GetOffset(uint32 bp, uint32 bw, uint32 psm);
-	Offset4* GetOffset4(const GIFRegFRAME& FRAME, const GIFRegZBUF& ZBUF);
+	BlockOffset* GetBlockOffset(uint32 bp, uint32 bw, uint32 psm);
+	PixelOffset* GetPixelOffset(uint32 bp, uint32 bw, uint32 psm);
+	PixelOffset4* GetPixelOffset4(const GIFRegFRAME& FRAME, const GIFRegZBUF& ZBUF);
 
 	// address
 
@@ -781,19 +790,19 @@ public:
 
 	// * => 32
 
-	void ReadTexture32(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture24(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture16(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture16S(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture8(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture8H(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4HL(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4HH(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture32Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture24Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture16Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture16SZ(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
+	void ReadTexture32(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture24(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture16(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture16S(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture8(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture8H(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4HL(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4HH(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture32Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture24Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture16Z(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture16SZ(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
 
 	void ReadTexture(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
 
@@ -813,25 +822,25 @@ public:
 
 	// * => 32/16
 
-	void ReadTexture16NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture16SNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture8NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture8HNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4HLNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4HHNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture16ZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture16SZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
+	void ReadTexture16NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture16SNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture8NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4NP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture8HNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4HLNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4HHNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture16ZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture16SZNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
 
 	void ReadTextureNP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
 
 	// pal ? 8 : 32
 
-	void ReadTexture8P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture8HP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4HLP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
-	void ReadTexture4HHP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA) const;
+	void ReadTexture8P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4P(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture8HP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4HLP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
+	void ReadTexture4HHP(const GSVector4i& r, uint8* dst, int dstpitch, const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
 
 	void ReadTextureBlock8P(uint32 bp, uint8* dst, int dstpitch, const GIFRegTEXA& TEXA) const;
 	void ReadTextureBlock4P(uint32 bp, uint8* dst, int dstpitch, const GIFRegTEXA& TEXA) const;

plugins/GSdx/GSRasterizer.cpp

@@ -290,8 +290,8 @@ void GSRasterizer::DrawTriangleTop(GSVertexSW* v, const GSVector4i& scissor)
 
 	GSVector4 tb = l.p.upl(v[2].p).ceil();
 
-	GSVector4 tbmax = tb.maxv(fscissor.yyyy());
-	GSVector4 tbmin = tb.minv(fscissor.wwww());
+	GSVector4 tbmax = tb.max(fscissor.yyyy());
+	GSVector4 tbmin = tb.min(fscissor.wwww());
 
 	GSVector4i tbi = GSVector4i(tbmax.zzww(tbmin));
 
@@ -342,8 +342,8 @@ void GSRasterizer::DrawTriangleBottom(GSVertexSW* v, const GSVector4i& scissor)
 
 	GSVector4 tb = l.p.upl(v[2].p).ceil();
 
-	GSVector4 tbmax = tb.maxv(fscissor.yyyy());
-	GSVector4 tbmin = tb.minv(fscissor.wwww());
+	GSVector4 tbmax = tb.max(fscissor.yyyy());
+	GSVector4 tbmin = tb.min(fscissor.wwww());
 
 	GSVector4i tbi = GSVector4i(tbmax.zzww(tbmin));
 
@@ -398,8 +398,8 @@ void GSRasterizer::DrawTriangleTopBottom(GSVertexSW* v, const GSVector4i& scisso
 
 	GSVector4 tb = v[0].p.upl(v[1].p).zwzw(v[1].p.upl(v[2].p)).ceil();
 
-	GSVector4 tbmax = tb.maxv(fscissor.yyyy());
-	GSVector4 tbmin = tb.minv(fscissor.wwww());
+	GSVector4 tbmax = tb.max(fscissor.yyyy());
+	GSVector4 tbmin = tb.min(fscissor.wwww());
 
 	GSVector4i tbi = GSVector4i(tbmax.xzyw(tbmin));
 
@@ -462,8 +462,8 @@ void GSRasterizer::DrawTriangleSection(int top, int bottom, GSVertexSW& l, const
 			{
 				GSVector4 lr = l.p.xyxy(r).ceil();
 
-				GSVector4 lrmax = lr.maxv(fscissor.xxxx());
-				GSVector4 lrmin = lr.minv(fscissor.zzzz());
+				GSVector4 lrmax = lr.max(fscissor.xxxx());
+				GSVector4 lrmin = lr.min(fscissor.zzzz());
 
 				GSVector4i lri = GSVector4i(lrmax.xxzz(lrmin));
 
@@ -503,8 +503,8 @@ void GSRasterizer::DrawTriangleSection(int top, int bottom, GSVertexSW& l, const
 			{
 				GSVector4 lr = l.p.ceil();
 
-				GSVector4 lrmax = lr.maxv(fscissor.xxxx());
-				GSVector4 lrmin = lr.minv(fscissor.zzzz());
+				GSVector4 lrmax = lr.max(fscissor.xxxx());
+				GSVector4 lrmin = lr.min(fscissor.zzzz());
 
 				GSVector4i lri = GSVector4i(lrmax.xxyy(lrmin));
 
@@ -613,8 +613,8 @@ void GSRasterizer::DrawEdge(const GSVertexSW& v0, const GSVertexSW& v1, const GS
 
 	if(orientation)
 	{
-		GSVector4 tbmax = lrtb.maxv(fscissor.yyyy());
-		GSVector4 tbmin = lrtb.minv(fscissor.wwww());
+		GSVector4 tbmax = lrtb.max(fscissor.yyyy());
+		GSVector4 tbmin = lrtb.min(fscissor.wwww());
 
 		GSVector4i tbi = GSVector4i(tbmax.zwzw(tbmin));
 
@@ -711,8 +711,8 @@ void GSRasterizer::DrawEdge(const GSVertexSW& v0, const GSVertexSW& v1, const GS
 	}
 	else
 	{
-		GSVector4 lrmax = lrtb.maxv(fscissor.xxxx());
-		GSVector4 lrmin = lrtb.minv(fscissor.zzzz());
+		GSVector4 lrmax = lrtb.max(fscissor.xxxx());
+		GSVector4 lrmin = lrtb.min(fscissor.zzzz());
 
 		GSVector4i lri = GSVector4i(lrmax.xyxy(lrmin));
 

plugins/GSdx/GSRenderer.cpp

@@ -259,9 +259,17 @@ void GSRenderer::VSync(int field)
 
 	Flush();
 
-	field = field ? 1 : 0;
-
-	if(!Merge(field)) return;
+	if(!m_dev->IsLost(true))
+	{
+		if(!Merge(field ? 1 : 0))
+		{
+			return;
+		}
+	}
+	else
+	{
+		ResetDevice();
+	}
 
 	// osd 
 
@@ -311,11 +319,6 @@ void GSRenderer::VSync(int field)
 
 	// present
 
-	if(m_dev->IsLost())
-	{
-		ResetDevice();
-	}
-
 	m_dev->Present(m_wnd.GetClientRect().fit(m_aspectratio), m_shader);
 
 	// snapshot

plugins/GSdx/GSRenderer.h

@@ -127,7 +127,10 @@ protected:
 			{
 				// FIXME: berserk fpsm = 27 (8H)
 
+				if(!m_dev->IsLost())
+				{
 					Draw();
+				}
 
 				m_perfmon.Put(GSPerfMon::Draw, 1);
 			}

plugins/GSdx/GSRendererDX9.cpp

@@ -122,14 +122,14 @@ void GSRendererDX9::VertexKick(bool skip)
 		case GS_LINELIST:
 		case GS_LINESTRIP:
 		case GS_SPRITE:
-			pmin = v[0].p.minv(v[1].p);
-			pmax = v[0].p.maxv(v[1].p);
+			pmin = v[0].p.min(v[1].p);
+			pmax = v[0].p.max(v[1].p);
 			break;
 		case GS_TRIANGLELIST:
 		case GS_TRIANGLESTRIP:
 		case GS_TRIANGLEFAN:
-			pmin = v[0].p.minv(v[1].p).minv(v[2].p);
-			pmax = v[0].p.maxv(v[1].p).maxv(v[2].p);
+			pmin = v[0].p.min(v[1].p).min(v[2].p);
+			pmax = v[0].p.max(v[1].p).max(v[2].p);
 			break;
 		}
 

plugins/GSdx/GSRendererSW.cpp

@@ -87,37 +87,19 @@ GSTexture* GSRendererSW::GetOutput(int i)
 	TEX0.TBW = DISPFB.FBW;
 	TEX0.PSM = DISPFB.PSM;
 
-	GSVector4i r(0, 0, TEX0.TBW * 64, GetFrameRect(i).bottom);
+	int w = TEX0.TBW * 64;
+	int h = GetFrameRect(i).bottom;
 
 	// TODO: round up bottom
 
-	int w = r.width();
-	int h = r.height();
-
-	if(m_texture[i])
+	if(m_dev->ResizeTexture(&m_texture[i], w, h))
 	{
-		if(m_texture[i]->GetWidth() != w || m_texture[i]->GetHeight() != h)
-		{
-			delete m_texture[i];
-
-			m_texture[i] = NULL;
-		}
-	}
-
-	if(!m_texture[i])
-	{
-		m_texture[i] = m_dev->CreateTexture(w, h);
-
-		if(!m_texture[i])
-		{
-			return NULL;
-		}
-	}
-
 		// TODO
 		static uint8* buff = (uint8*)_aligned_malloc(1024 * 1024 * 4, 16);
 		static int pitch = 1024 * 4;
 
+		GSVector4i r(0, 0, w, h);
+
 		m_mem.ReadTexture(r, buff, pitch, TEX0, m_env.TEXA);
 
 		m_texture[i]->Update(r, buff, pitch);
@@ -131,6 +113,7 @@ GSTexture* GSRendererSW::GetOutput(int i)
 
 			s_n++;
 		}
+	}
 
 	return m_texture[i];
 }
@@ -273,9 +256,9 @@ void GSRendererSW::GetScanlineParam(GSScanlineParam& p, GS_PRIM_CLASS primclass)
 
 	p.vm = m_mem.m_vm8;
 
-	p.fbo = m_mem.GetOffset(context->FRAME.Block(), context->FRAME.FBW, context->FRAME.PSM);
-	p.zbo = m_mem.GetOffset(context->ZBUF.Block(), context->FRAME.FBW, context->ZBUF.PSM);
-	p.fzbo = m_mem.GetOffset4(context->FRAME, context->ZBUF);
+	p.fbo = m_mem.GetPixelOffset(context->FRAME.Block(), context->FRAME.FBW, context->FRAME.PSM);
+	p.zbo = m_mem.GetPixelOffset(context->ZBUF.Block(), context->FRAME.FBW, context->ZBUF.PSM);
+	p.fzbo = m_mem.GetPixelOffset4(context->FRAME, context->ZBUF);
 
 	p.sel.key = 0;
 
@@ -517,14 +500,14 @@ if(!m_dump)
 		case GS_LINELIST:
 		case GS_LINESTRIP:
 		case GS_SPRITE:
-			pmin = v[0].p.minv(v[1].p);
-			pmax = v[0].p.maxv(v[1].p);
+			pmin = v[0].p.min(v[1].p);
+			pmax = v[0].p.max(v[1].p);
 			break;
 		case GS_TRIANGLELIST:
 		case GS_TRIANGLESTRIP:
 		case GS_TRIANGLEFAN:
-			pmin = v[0].p.minv(v[1].p).minv(v[2].p);
-			pmax = v[0].p.maxv(v[1].p).maxv(v[2].p);
+			pmin = v[0].p.min(v[1].p).min(v[2].p);
+			pmax = v[0].p.max(v[1].p).max(v[2].p);
 			break;
 		}
 

plugins/GSdx/GSScanlineEnvironment.h

@@ -108,9 +108,9 @@ __declspec(align(16)) struct GSScanlineParam
 	const uint32* clut;
 	uint32 tw;
 
-	GSLocalMemory::Offset* fbo;
-	GSLocalMemory::Offset* zbo;
-	GSLocalMemory::Offset4* fzbo;
+	GSLocalMemory::PixelOffset* fbo;
+	GSLocalMemory::PixelOffset* zbo;
+	GSLocalMemory::PixelOffset4* fzbo;
 
 	uint32 fm, zm;
 };
@@ -122,10 +122,10 @@ __declspec(align(16)) struct GSScanlineEnvironment
 	const uint32* clut;
 	uint32 tw;
 
-	GSVector4i* fbr;
-	GSVector4i* zbr;
-	int** fbc;
-	int** zbc;
+	int* fbr;
+	int* zbr;
+	int* fbc;
+	int* zbc;
 	GSVector2i* fzbr;
 	GSVector2i* fzbc;
 

plugins/GSdx/GSState.cpp

@@ -948,7 +948,7 @@ void GSState::FlushWrite()
 	r.left = m_env.TRXPOS.DSAX;
 	r.top = y;
 	r.right = r.left + m_env.TRXREG.RRW;
-	r.bottom = min(r.top + m_env.TRXREG.RRH, m_tr.x == r.left ? m_tr.y : m_tr.y + 1);
+	r.bottom = std::min<int>(r.top + m_env.TRXREG.RRH, m_tr.x == r.left ? m_tr.y : m_tr.y + 1);
 
 	InvalidateVideoMem(m_env.BITBLTBUF, r);
 /*
@@ -1085,17 +1085,20 @@ void GSState::Move()
 
 	// TODO: unroll inner loops (width has special size requirement, must be multiples of 1 << n, depending on the format)
 
+	GSLocalMemory::PixelOffset* RESTRICT spo = m_mem.GetPixelOffset(m_env.BITBLTBUF.SBP, m_env.BITBLTBUF.SBW, m_env.BITBLTBUF.SPSM);
+	GSLocalMemory::PixelOffset* RESTRICT dpo = m_mem.GetPixelOffset(m_env.BITBLTBUF.DBP, m_env.BITBLTBUF.DBW, m_env.BITBLTBUF.DPSM);
+
 	if(spsm.trbpp == dpsm.trbpp && spsm.trbpp >= 16)
 	{
-		int* soffset = spsm.rowOffset[0];
-		int* doffset = dpsm.rowOffset[0];
+		int* soffset = spo->col[0];
+		int* doffset = dpo->col[0];
 
 		if(spsm.trbpp == 32)
 		{
 			for(int y = 0; y < h; y++, sy += yinc, dy += yinc, sx -= xinc * w, dx -= xinc * w)
 			{
-				uint32 sbase = spsm.pa(0, sy, m_env.BITBLTBUF.SBP, m_env.BITBLTBUF.SBW);
-				uint32 dbase = dpsm.pa(0, dy, m_env.BITBLTBUF.DBP, m_env.BITBLTBUF.DBW);
+				uint32 sbase = spo->row[sy];
+				uint32 dbase = dpo->row[dy];
 				
 				for(int x = 0; x < w; x++, sx += xinc, dx += xinc)
 				{
@@ -1107,8 +1110,8 @@ void GSState::Move()
 		{
 			for(int y = 0; y < h; y++, sy += yinc, dy += yinc, sx -= xinc * w, dx -= xinc * w)
 			{
-				uint32 sbase = spsm.pa(0, sy, m_env.BITBLTBUF.SBP, m_env.BITBLTBUF.SBW);
-				uint32 dbase = dpsm.pa(0, dy, m_env.BITBLTBUF.DBP, m_env.BITBLTBUF.DBW);
+				uint32 sbase = spo->row[sy];
+				uint32 dbase = dpo->row[dy];
 				
 				for(int x = 0; x < w; x++, sx += xinc, dx += xinc)
 				{
@@ -1120,8 +1123,8 @@ void GSState::Move()
 		{
 			for(int y = 0; y < h; y++, sy += yinc, dy += yinc, sx -= xinc * w, dx -= xinc * w)
 			{
-				uint32 sbase = spsm.pa(0, sy, m_env.BITBLTBUF.SBP, m_env.BITBLTBUF.SBW);
-				uint32 dbase = dpsm.pa(0, dy, m_env.BITBLTBUF.DBP, m_env.BITBLTBUF.DBW);
+				uint32 sbase = spo->row[sy];
+				uint32 dbase = dpo->row[dy];
 				
 				for(int x = 0; x < w; x++, sx += xinc, dx += xinc)
 				{
@@ -1134,11 +1137,11 @@ void GSState::Move()
 	{
 		for(int y = 0; y < h; y++, sy += yinc, dy += yinc, sx -= xinc * w, dx -= xinc * w)
 		{
-			uint32 sbase = GSLocalMemory::PixelAddress8(0, sy, m_env.BITBLTBUF.SBP, m_env.BITBLTBUF.SBW);
-			int* soffset = spsm.rowOffset[sy & 7];
+			uint32 sbase = spo->row[sy];
+			uint32 dbase = dpo->row[dy];
 
-			uint32 dbase = GSLocalMemory::PixelAddress8(0, dy, m_env.BITBLTBUF.DBP, m_env.BITBLTBUF.DBW);
-			int* doffset = dpsm.rowOffset[dy & 7];
+			int* soffset = spo->col[sy & 7];
+			int* doffset = dpo->col[dy & 7];
 			
 			for(int x = 0; x < w; x++, sx += xinc, dx += xinc)
 			{
@@ -1150,11 +1153,11 @@ void GSState::Move()
 	{
 		for(int y = 0; y < h; y++, sy += yinc, dy += yinc, sx -= xinc * w, dx -= xinc * w)
 		{
-			uint32 sbase = GSLocalMemory::PixelAddress4(0, sy, m_env.BITBLTBUF.SBP, m_env.BITBLTBUF.SBW);
-			int* soffset = spsm.rowOffset[sy & 7];
+			uint32 sbase = spo->row[sy];
+			uint32 dbase = dpo->row[dy];
 
-			uint32 dbase = GSLocalMemory::PixelAddress4(0, dy, m_env.BITBLTBUF.DBP, m_env.BITBLTBUF.DBW);
-			int* doffset = dpsm.rowOffset[dy & 7];
+			int* soffset = spo->col[sy & 7];
+			int* doffset = dpo->col[dy & 7];
 			
 			for(int x = 0; x < w; x++, sx += xinc, dx += xinc)
 			{
@@ -1166,11 +1169,11 @@ void GSState::Move()
 	{
 		for(int y = 0; y < h; y++, sy += yinc, dy += yinc, sx -= xinc * w, dx -= xinc * w)
 		{
-			uint32 sbase = spsm.pa(0, sy, m_env.BITBLTBUF.SBP, m_env.BITBLTBUF.SBW);
-			int* soffset = spsm.rowOffset[sy & 7];
+			uint32 sbase = spo->row[sy];
+			uint32 dbase = dpo->row[dy];
 
-			uint32 dbase = dpsm.pa(0, dy, m_env.BITBLTBUF.DBP, m_env.BITBLTBUF.DBW);
-			int* doffset = dpsm.rowOffset[dy & 7];
+			int* soffset = spo->col[sy & 7];
+			int* doffset = dpo->col[dy & 7];
 			
 			for(int x = 0; x < w; x++, sx += xinc, dx += xinc)
 			{

plugins/GSdx/GSTexture7.cpp

@@ -82,7 +82,7 @@ bool GSTexture7::Update(const GSVector4i& r, const void* data, int pitch)
 		uint8* src = (uint8*)data;
 		uint8* dst = (uint8*)desc.lpSurface;
 
-		int bytes = min(pitch, desc.lPitch);
+		int bytes = std::min<int>(pitch, desc.lPitch);
 
 		for(int i = 0, j = r.height(); i < j; i++, src += pitch, dst += desc.lPitch)
 		{

plugins/GSdx/GSTextureCache.cpp

@@ -60,7 +60,7 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const GIFRegTEX0& TEX0, con
 	{
 		Source* s = i->first;
 
-		if(((s->m_TEX0.u32[0] ^ TEX0.u32[0]) | ((s->m_TEX0.u32[1] ^ TEX0.u32[1]) & 3)) != 0) // TBP0 TBW PSM TW TH
+		if(((TEX0.u32[0] ^ s->m_TEX0.u32[0]) | ((TEX0.u32[1] ^ s->m_TEX0.u32[1]) & 3)) != 0) // TBP0 TBW PSM TW TH
 		{
 			continue;
 		}
@@ -70,7 +70,7 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const GIFRegTEX0& TEX0, con
 			continue;
 		}
 
-		if(psm.pal > 0 && !GSVector4i::compare(s->m_clut, clut, psm.pal * sizeof(clut[0])))
+		if(psm.pal > 0 && !GSVector4i::compare(clut, s->m_clut, psm.pal * sizeof(clut[0])))
 		{
 			continue;
 		}
@@ -84,13 +84,16 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const GIFRegTEX0& TEX0, con
 
 	if(src == NULL)
 	{
+		uint32 bp = TEX0.TBP0;
+		uint32 psm = TEX0.PSM;
+
 		for(int type = 0; type < 2 && dst == NULL; type++)
 		{
 			for(list<Target*>::iterator i = m_dst[type].begin(); i != m_dst[type].end(); i++)
 			{
 				Target* t = *i;
 
-				if(t->m_used && t->m_dirty.empty() && GSUtil::HasSharedBits(t->m_TEX0.TBP0, t->m_TEX0.PSM, TEX0.TBP0, TEX0.PSM))
+				if(t->m_used && t->m_dirty.empty() && GSUtil::HasSharedBits(bp, psm, t->m_TEX0.TBP0, t->m_TEX0.PSM))
 				{
 					dst = t;
 
@@ -116,7 +119,7 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const GIFRegTEX0& TEX0, con
 			memcpy(src->m_clut, clut, psm.pal * sizeof(clut[0]));
 		}
 
-		m_src.Add(src, TEX0);
+		m_src.Add(src, TEX0, m_renderer->m_mem);
 	}
 
 	if(psm.pal > 0)
@@ -142,13 +145,15 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const GIFRegTEX0& TEX0, con
 
 GSTextureCache::Target* GSTextureCache::LookupTarget(const GIFRegTEX0& TEX0, int w, int h, int type, bool used, bool fb)
 {
+	uint32 bp = TEX0.TBP0;
+
 	Target* dst = NULL;
 
 	for(list<Target*>::iterator i = m_dst[type].begin(); i != m_dst[type].end(); i++)
 	{
 		Target* t = *i;
 
-		if(t->m_TEX0.TBP0 == TEX0.TBP0)
+		if(bp == t->m_TEX0.TBP0)
 		{
 			m_dst[type].splice(m_dst[type].begin(), m_dst[type], i);
 
@@ -168,7 +173,7 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(const GIFRegTEX0& TEX0, int
 		{
 			Target* t = *i;
 
-			if(t->m_TEX0.TBP0 <= TEX0.TBP0 && TEX0.TBP0 < t->m_TEX0.TBP0 + 0x700 && (!dst || t->m_TEX0.TBP0 >= dst->m_TEX0.TBP0))
+			if(t->m_TEX0.TBP0 <= bp && bp < t->m_TEX0.TBP0 + 0x700 && (!dst || t->m_TEX0.TBP0 >= dst->m_TEX0.TBP0))
 			{
 				dst = t;
 			}
@@ -229,14 +234,13 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(const GIFRegTEX0& TEX0, int
 
 void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& rect, bool target)
 {
-	bool found = false;
-
-	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[BITBLTBUF.DPSM];
-
 	uint32 bp = BITBLTBUF.DBP;
 	uint32 bw = BITBLTBUF.DBW;
+	uint32 psm = BITBLTBUF.DPSM;
 
-	GSVector2i bs = (bp & 31) == 0 ? psm.pgs : psm.bs;
+	const GSLocalMemory::BlockOffset* bo = m_renderer->m_mem.GetBlockOffset(bp, bw, psm);
+
+	GSVector2i bs = (bp & 31) == 0 ? GSLocalMemory::m_psm[psm].pgs : GSLocalMemory::m_psm[psm].bs;
 
 	GSVector4i r = rect.ralign<GSVector4i::Outside>(bs);
 
@@ -250,20 +254,22 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 
 			Source* s = j->first;
 
-			if(GSUtil::HasSharedBits(bp, BITBLTBUF.DPSM, s->m_TEX0.TBP0, s->m_TEX0.PSM))
+			if(GSUtil::HasSharedBits(bp, psm, s->m_TEX0.TBP0, s->m_TEX0.PSM))
 			{
 				m_src.RemoveAt(s);
 			}
 		}
 	}
 
+	bool found = false;
+
 	for(int y = r.top; y < r.bottom; y += bs.y)
 	{
-		uint32 base = psm.bn(0, y, bp, bw);
+		uint32 base = bo->row[y >> 3];
 
 		for(int x = r.left; x < r.right; x += bs.x)
 		{
-			uint32 page = (base + psm.blockOffset[x >> 3]) >> 5;
+			uint32 page = (base + bo->col[x >> 3]) >> 5;
 
 			if(page < MAX_PAGES)
 			{
@@ -275,20 +281,22 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 
 					Source* s = j->first;
 
-					if(GSUtil::HasSharedBits(BITBLTBUF.DPSM, s->m_TEX0.PSM))
+					if(GSUtil::HasSharedBits(psm, s->m_TEX0.PSM))
 					{
+						bool b = bp == s->m_TEX0.TBP0;
+
 						if(!s->m_target)
 						{
 							s->m_valid[page] = 0;
 							s->m_complete = false;
 
-							found = true;
+							found = b;
 						}
 						else
 						{
 							// TODO
 
-							if(s->m_TEX0.TBP0 == bp)
+							if(b)
 							{
 								m_src.RemoveAt(s);
 							}
@@ -309,12 +317,12 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 
 			Target* t = *j;
 
-			if(GSUtil::HasSharedBits(BITBLTBUF.DBP, BITBLTBUF.DPSM, t->m_TEX0.TBP0, t->m_TEX0.PSM))
+			if(GSUtil::HasSharedBits(bp, psm, t->m_TEX0.TBP0, t->m_TEX0.PSM))
 			{
-				if(!found && GSUtil::HasCompatibleBits(BITBLTBUF.DPSM, t->m_TEX0.PSM))
+				if(!found && GSUtil::HasCompatibleBits(psm, t->m_TEX0.PSM))
 				{
-					t->m_dirty.push_back(GSDirtyRect(r, BITBLTBUF.DPSM));
-					t->m_TEX0.TBW = BITBLTBUF.DBW;
+					t->m_dirty.push_back(GSDirtyRect(r, psm));
+					t->m_TEX0.TBW = bw;
 				}
 				else
 				{
@@ -324,20 +332,20 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 				}
 			}
 
-			if(GSUtil::HasSharedBits(BITBLTBUF.DPSM, t->m_TEX0.PSM) && BITBLTBUF.DBP < t->m_TEX0.TBP0)
+			if(GSUtil::HasSharedBits(psm, t->m_TEX0.PSM) && bp < t->m_TEX0.TBP0)
 			{
-				uint32 rowsize = BITBLTBUF.DBW * 8192;
-				uint32 offset = (uint32)((t->m_TEX0.TBP0 - BITBLTBUF.DBP) * 256);
+				uint32 rowsize = bw * 8192;
+				uint32 offset = (uint32)((t->m_TEX0.TBP0 - bp) * 256);
 
 				if(rowsize > 0 && offset % rowsize == 0)
 				{
-					int y = GSLocalMemory::m_psm[BITBLTBUF.DPSM].pgs.y * offset / rowsize;
+					int y = GSLocalMemory::m_psm[psm].pgs.y * offset / rowsize;
 
 					if(r.bottom > y)
 					{
 						// TODO: do not add this rect above too
-						t->m_dirty.push_back(GSDirtyRect(GSVector4i(r.left, r.top - y, r.right, r.bottom - y), BITBLTBUF.DPSM));
-						t->m_TEX0.TBW = BITBLTBUF.DBW;
+						t->m_dirty.push_back(GSDirtyRect(GSVector4i(r.left, r.top - y, r.right, r.bottom - y), psm));
+						t->m_TEX0.TBW = bw;
 						continue;
 					}
 				}
@@ -348,21 +356,24 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 
 void GSTextureCache::InvalidateLocalMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r)
 {
+	uint32 bp = BITBLTBUF.SBP;
+	uint32 psm = BITBLTBUF.SPSM;
+
 	for(list<Target*>::iterator i = m_dst[RenderTarget].begin(); i != m_dst[RenderTarget].end(); )
 	{
 		list<Target*>::iterator j = i++;
 
 		Target* t = *j;
 
-		if(GSUtil::HasSharedBits(BITBLTBUF.SBP, BITBLTBUF.SPSM, t->m_TEX0.TBP0, t->m_TEX0.PSM))
+		if(GSUtil::HasSharedBits(bp, psm, t->m_TEX0.TBP0, t->m_TEX0.PSM))
 		{
-			if(GSUtil::HasCompatibleBits(BITBLTBUF.SPSM, t->m_TEX0.PSM))
+			if(GSUtil::HasCompatibleBits(psm, t->m_TEX0.PSM))
 			{
 				t->Read(r);
 
 				return;
 			}
-			else if(BITBLTBUF.SPSM == PSM_PSMCT32 && (t->m_TEX0.PSM == PSM_PSMCT16 || t->m_TEX0.PSM == PSM_PSMCT16S)) 
+			else if(psm == PSM_PSMCT32 && (t->m_TEX0.PSM == PSM_PSMCT16 || t->m_TEX0.PSM == PSM_PSMCT16S)) 
 			{
 				// ffx-2 riku changing to her default (shoots some reflecting glass at the end), 16-bit rt read as 32-bit
 
@@ -518,9 +529,7 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 	m_TEX0 = TEX0;
 	m_TEXA = TEXA;
 
-	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[m_TEX0.PSM];
-
-	GSVector2i bs = psm.bs;
+	GSVector2i bs = GSLocalMemory::m_psm[m_TEX0.PSM].bs;
 
 	GSVector4i r = rect.ralign<GSVector4i::Outside>(bs);
 
@@ -529,20 +538,19 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 		m_complete = true; // lame, but better than nothing
 	}
 
-	uint32 bp = m_TEX0.TBP0;
-	uint32 bw = m_TEX0.TBW;
+	const GSLocalMemory::BlockOffset* bo = m_renderer->m_mem.GetBlockOffset(m_TEX0.TBP0, m_TEX0.TBW, m_TEX0.PSM);
 
-	bool repeating = (1 << m_TEX0.TW) > (bw << 6); // TODO: bw == 0
+	bool repeating = m_TEX0.IsRepeating();
 
 	uint32 blocks = 0;
 
 	for(int y = r.top; y < r.bottom; y += bs.y)
 	{
-		uint32 base = psm.bn(0, y, bp, bw);
+		uint32 base = bo->row[y >> 3];
 
 		for(int x = r.left; x < r.right; x += bs.x)
 		{
-			uint32 block = base + psm.blockOffset[x >> 3];
+			uint32 block = base + bo->col[x >> 3];
 
 			if(block < MAX_BLOCKS)
 			{
@@ -570,11 +578,11 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 		{
 			for(int y = r.top; y < r.bottom; y += bs.y)
 			{
-				uint32 base = psm.bn(0, y, bp, bw);
+				uint32 base = bo->row[y >> 3];
 
 				for(int x = r.left; x < r.right; x += bs.x)
 				{
-					uint32 block = base + psm.blockOffset[x >> 3];
+					uint32 block = base + bo->col[x >> 3];
 
 					if(block < MAX_BLOCKS)
 					{
@@ -640,7 +648,7 @@ void GSTextureCache::Source::Flush(uint32 count)
 	
 	int pitch = max(tw, psm.bs.x) * sizeof(uint32);
 
-	const GSLocalMemory& mem = m_renderer->m_mem;
+	GSLocalMemory& mem = m_renderer->m_mem;
 
 	GSLocalMemory::readTexture rtx = psm.rtx;
 
@@ -767,42 +775,44 @@ void GSTextureCache::Target::Update()
 	{
 		// do the most likely thing a direct write would do, clear it
 
+		if((m_renderer->m_game.flags & CRC::ZWriteMustNotClear) == 0)
+		{
 			m_renderer->m_dev->ClearDepth(m_texture, 0);
 		}
 	}
+}
 
 // GSTextureCache::SourceMap
 
-void GSTextureCache::SourceMap::Add(Source* s, const GIFRegTEX0& TEX0)
+void GSTextureCache::SourceMap::Add(Source* s, const GIFRegTEX0& TEX0, GSLocalMemory& mem)
 {
 	m_surfaces[s] = true;
 
-	int tw = 1 << TEX0.TW;
-	int th = 1 << TEX0.TH;
-
-	uint32 bp = TEX0.TBP0;
-	uint32 bw = TEX0.TBW;
-
 	if(s->m_target)
 	{
 		// TODO
 
-		m_map[bp >> 5][s] = true;
+		m_map[TEX0.TBP0 >> 5][s] = true;
 
 		return;
 	}
 
+	const GSLocalMemory::BlockOffset* bo = mem.GetBlockOffset(TEX0.TBP0, TEX0.TBW, TEX0.PSM);
+
 	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[TEX0.PSM];
 
-	GSVector2i bs = (bp & 31) ? psm.pgs : psm.bs;
+	GSVector2i bs = (TEX0.TBP0 & 31) ? psm.pgs : psm.bs;
+
+	int tw = 1 << TEX0.TW;
+	int th = 1 << TEX0.TH;
 
 	for(int y = 0; y < th; y += bs.y)
 	{
-		uint32 base = psm.bn(0, y, bp, bw);
+		uint32 base = bo->row[y >> 3];
 
 		for(int x = 0; x < tw; x += bs.x)
 		{
-			uint32 page = (base + psm.blockOffset[x >> 3]) >> 5;
+			uint32 page = (base + bo->col[x >> 3]) >> 5;
 
 			if(page < MAX_PAGES)
 			{

plugins/GSdx/GSTextureCache.h

@@ -100,7 +100,7 @@ protected:
 
 		SourceMap() : m_used(false) {memset(m_pages, 0, sizeof(m_pages));}
 
-		void Add(Source* s, const GIFRegTEX0& TEX0);
+		void Add(Source* s, const GIFRegTEX0& TEX0, GSLocalMemory& mem);
 		void RemoveAll();
 		void RemoveAt(Source* s);
 

plugins/GSdx/GSTextureCacheSW.cpp

@@ -45,7 +45,7 @@ const GSTextureCacheSW::GSTexture* GSTextureCacheSW::Lookup(const GIFRegTEX0& TE
 	{
 		GSTexture* t2 = i->first;
 
-		if(((t2->m_TEX0.u32[0] ^ TEX0.u32[0]) | ((t2->m_TEX0.u32[1] ^ TEX0.u32[1]) & 3)) != 0) // TBP0 TBW PSM TW TH
+		if(((TEX0.u32[0] ^ t2->m_TEX0.u32[0]) | ((TEX0.u32[1] ^ t2->m_TEX0.u32[1]) & 3)) != 0) // TBP0 TBW PSM TW TH
 		{
 			continue;
 		}
@@ -68,21 +68,20 @@ const GSTextureCacheSW::GSTexture* GSTextureCacheSW::Lookup(const GIFRegTEX0& TE
 
 		m_textures[t] = true;
 
+		const GSLocalMemory::BlockOffset* bo = m_state->m_mem.GetBlockOffset(TEX0.TBP0, TEX0.TBW, TEX0.PSM);
+
+		GSVector2i bs = (TEX0.TBP0 & 31) == 0 ? psm.pgs : psm.bs;
+
 		int tw = 1 << TEX0.TW;
 		int th = 1 << TEX0.TH;
 
-		uint32 bp = TEX0.TBP0;
-		uint32 bw = TEX0.TBW;
-
-		GSVector2i s = (bp & 31) == 0 ? psm.pgs : psm.bs;
-
-		for(int y = 0; y < th; y += s.y)
+		for(int y = 0; y < th; y += bs.y)
 		{
-			uint32 base = psm.bn(0, y, bp, bw);
+			uint32 base = bo->row[y >> 3];
 
-			for(int x = 0; x < tw; x += s.x)
+			for(int x = 0; x < tw; x += bs.x)
 			{
-				uint32 page = (base + psm.blockOffset[x >> 3]) >> 5;
+				uint32 page = (base + bo->col[x >> 3]) >> 5;
 
 				if(page < MAX_PAGES)
 				{
@@ -168,22 +167,23 @@ void GSTextureCacheSW::IncAge()
 
 void GSTextureCacheSW::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& rect)
 {
-	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[BITBLTBUF.DPSM];
-
 	uint32 bp = BITBLTBUF.DBP;
 	uint32 bw = BITBLTBUF.DBW;
+	uint32 psm = BITBLTBUF.DPSM;
 
-	GSVector2i s = (bp & 31) == 0 ? psm.pgs : psm.bs;
+	const GSLocalMemory::BlockOffset* bo = m_state->m_mem.GetBlockOffset(bp, bw, psm);
 
-	GSVector4i r = rect.ralign<GSVector4i::Outside>(s);
+	GSVector2i bs = (bp & 31) == 0 ? GSLocalMemory::m_psm[psm].pgs : GSLocalMemory::m_psm[psm].bs;
 
-	for(int y = r.top; y < r.bottom; y += s.y)
+	GSVector4i r = rect.ralign<GSVector4i::Outside>(bs);
+
+	for(int y = r.top; y < r.bottom; y += bs.y)
 	{
-		uint32 base = psm.bn(0, y, bp, bw);
+		uint32 base = bo->row[y >> 3];
 
-		for(int x = r.left; x < r.right; x += s.x)
+		for(int x = r.left; x < r.right; x += bs.x)
 		{
-			uint32 page = (base + psm.blockOffset[x >> 3]) >> 5;
+			uint32 page = (base + bo->col[x >> 3]) >> 5;
 
 			if(page < MAX_PAGES)
 			{
@@ -193,7 +193,7 @@ void GSTextureCacheSW::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, cons
 				{
 					GSTexture* t = i->first;
 
-					if(GSUtil::HasSharedBits(BITBLTBUF.DPSM, t->m_TEX0.PSM))
+					if(GSUtil::HasSharedBits(psm, t->m_TEX0.PSM))
 					{
 						t->m_valid[page] = 0;
 						t->m_complete = false;
@@ -236,10 +236,17 @@ bool GSTextureCacheSW::GSTexture::Update(const GIFRegTEX0& TEX0, const GIFRegTEX
 
 	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[TEX0.PSM];
 
-	GSVector2i s = psm.bs;
+	GSVector2i bs = psm.bs;
 
-	int tw = max(1 << TEX0.TW, s.x);
-	int th = max(1 << TEX0.TH, s.y);
+	int tw = max(1 << TEX0.TW, bs.x);
+	int th = max(1 << TEX0.TH, bs.y);
+
+	GSVector4i r = rect.ralign<GSVector4i::Outside>(bs);
+
+	if(r.eq(GSVector4i(0, 0, tw, th)))
+	{
+		m_complete = true; // lame, but better than nothing
+	}
 
 	if(m_buff == NULL)
 	{
@@ -250,25 +257,17 @@ bool GSTextureCacheSW::GSTexture::Update(const GIFRegTEX0& TEX0, const GIFRegTEX
 			return false;
 		}
 
-		m_tw = max(psm.pal > 0 ? 5 : 3, TEX0.TW); // makes one row 32 bytes at least, matches the smallest block size that is allocated above for m_buff
+		m_tw = std::max<int>(psm.pal > 0 ? 5 : 3, TEX0.TW); // makes one row 32 bytes at least, matches the smallest block size that is allocated above for m_buff
 	}
 
-	GSVector4i r = rect.ralign<GSVector4i::Outside>(s);
-
-	if(r.eq(GSVector4i(0, 0, tw, th)))
-	{
-		m_complete = true; // lame, but better than nothing
-	}
-
-	uint32 bp = TEX0.TBP0;
-	uint32 bw = TEX0.TBW;
-
-	bool repeating = tw > (bw << 6); // TODO: bw == 0
-
-	uint32 blocks = 0;
-
 	GSLocalMemory& mem = m_state->m_mem;
 
+	const GSLocalMemory::BlockOffset* bo = mem.GetBlockOffset(m_TEX0.TBP0, m_TEX0.TBW, m_TEX0.PSM);
+
+	bool repeating = m_TEX0.IsRepeating();
+
+	uint32 blocks = 0;
+
 	GSLocalMemory::readTextureBlock rtxb = psm.rtxbP;
 
 	int shift = psm.pal == 0 ? 2 : 0;
@@ -277,13 +276,13 @@ bool GSTextureCacheSW::GSTexture::Update(const GIFRegTEX0& TEX0, const GIFRegTEX
 
 	uint8* dst = (uint8*)m_buff + pitch * r.top;
 
-	for(int y = r.top, o = pitch * s.y; y < r.bottom; y += s.y, dst += o)
+	for(int y = r.top, o = pitch * bs.y; y < r.bottom; y += bs.y, dst += o)
 	{
-		uint32 base = psm.bn(0, y, bp, bw);
+		uint32 base = bo->row[y >> 3];
 
-		for(int x = r.left; x < r.right; x += s.x)
+		for(int x = r.left; x < r.right; x += bs.x)
 		{
-			uint32 block = base + psm.blockOffset[x >> 3];
+			uint32 block = base + bo->col[x >> 3];
 
 			if(block < MAX_BLOCKS)
 			{
@@ -309,13 +308,13 @@ bool GSTextureCacheSW::GSTexture::Update(const GIFRegTEX0& TEX0, const GIFRegTEX
 	{
 		if(repeating)
 		{
-			for(int y = r.top; y < r.bottom; y += s.y)
+			for(int y = r.top; y < r.bottom; y += bs.y)
 			{
-				uint32 base = psm.bn(0, y, bp, bw);
+				uint32 base = bo->row[y >> 3];
 
-				for(int x = r.left; x < r.right; x += s.x)
+				for(int x = r.left; x < r.right; x += bs.x)
 				{
-					uint32 block = base + psm.blockOffset[x >> 3];
+					uint32 block = base + bo->col[x >> 3];
 
 					if(block < MAX_BLOCKS)
 					{
@@ -328,7 +327,7 @@ bool GSTextureCacheSW::GSTexture::Update(const GIFRegTEX0& TEX0, const GIFRegTEX
 			}
 		}
 
-		m_state->m_perfmon.Put(GSPerfMon::Unswizzle, s.x * s.y * blocks << shift);
+		m_state->m_perfmon.Put(GSPerfMon::Unswizzle, bs.x * bs.y * blocks << shift);
 	}
 
 	return true;

plugins/GSdx/GSUtil.cpp

@@ -28,8 +28,8 @@
 static struct GSUtilMaps
 {
 	uint8 PrimClassField[8];
-	bool CompatibleBitsField[64][64];
-	bool SharedBitsField[64][64];
+	uint32 CompatibleBitsField[64][2];
+	uint32 SharedBitsField[64][2];
 
 	struct GSUtilMaps()
 	{
@@ -44,31 +44,36 @@ static struct GSUtilMaps
 
 		memset(CompatibleBitsField, 0, sizeof(CompatibleBitsField));
 
-		CompatibleBitsField[PSM_PSMCT32][PSM_PSMCT24] = true;
-		CompatibleBitsField[PSM_PSMCT24][PSM_PSMCT32] = true;
-		CompatibleBitsField[PSM_PSMCT16][PSM_PSMCT16S] = true;
-		CompatibleBitsField[PSM_PSMCT16S][PSM_PSMCT16] = true;
-		CompatibleBitsField[PSM_PSMZ32][PSM_PSMZ24] = true;
-		CompatibleBitsField[PSM_PSMZ24][PSM_PSMZ32] = true;
-		CompatibleBitsField[PSM_PSMZ16][PSM_PSMZ16S] = true;
-		CompatibleBitsField[PSM_PSMZ16S][PSM_PSMZ16] = true;
+		for(int i = 0; i < 64; i++)
+		{
+			CompatibleBitsField[i][i >> 5] |= 1 << (i & 0x1f);
+		}
 
-		memset(SharedBitsField, 1, sizeof(SharedBitsField));
+		CompatibleBitsField[PSM_PSMCT32][PSM_PSMCT24 >> 5] |= 1 << (PSM_PSMCT24 & 0x1f);
+		CompatibleBitsField[PSM_PSMCT24][PSM_PSMCT32 >> 5] |= 1 << (PSM_PSMCT32 & 0x1f);
+		CompatibleBitsField[PSM_PSMCT16][PSM_PSMCT16S >> 5] |= 1 << (PSM_PSMCT16S & 0x1f);
+		CompatibleBitsField[PSM_PSMCT16S][PSM_PSMCT16 >> 5] |= 1 << (PSM_PSMCT16 & 0x1f);
+		CompatibleBitsField[PSM_PSMZ32][PSM_PSMZ24 >> 5] |= 1 << (PSM_PSMZ24 & 0x1f);
+		CompatibleBitsField[PSM_PSMZ24][PSM_PSMZ32 >> 5] |= 1 << (PSM_PSMZ32 & 0x1f);
+		CompatibleBitsField[PSM_PSMZ16][PSM_PSMZ16S >> 5] |= 1 << (PSM_PSMZ16S & 0x1f);
+		CompatibleBitsField[PSM_PSMZ16S][PSM_PSMZ16 >> 5] |= 1 << (PSM_PSMZ16 & 0x1f);
 
-		SharedBitsField[PSM_PSMCT24][PSM_PSMT8H] = false;
-		SharedBitsField[PSM_PSMCT24][PSM_PSMT4HL] = false;
-		SharedBitsField[PSM_PSMCT24][PSM_PSMT4HH] = false;
-		SharedBitsField[PSM_PSMZ24][PSM_PSMT8H] = false;
-		SharedBitsField[PSM_PSMZ24][PSM_PSMT4HL] = false;
-		SharedBitsField[PSM_PSMZ24][PSM_PSMT4HH] = false;
-		SharedBitsField[PSM_PSMT8H][PSM_PSMCT24] = false;
-		SharedBitsField[PSM_PSMT8H][PSM_PSMZ24] = false;
-		SharedBitsField[PSM_PSMT4HL][PSM_PSMCT24] = false;
-		SharedBitsField[PSM_PSMT4HL][PSM_PSMZ24] = false;
-		SharedBitsField[PSM_PSMT4HL][PSM_PSMT4HH] = false;
-		SharedBitsField[PSM_PSMT4HH][PSM_PSMCT24] = false;
-		SharedBitsField[PSM_PSMT4HH][PSM_PSMZ24] = false;
-		SharedBitsField[PSM_PSMT4HH][PSM_PSMT4HL] = false;
+		memset(SharedBitsField, 0, sizeof(SharedBitsField));
+
+		SharedBitsField[PSM_PSMCT24][PSM_PSMT8H >> 5] |= 1 << (PSM_PSMT8H & 0x1f);
+		SharedBitsField[PSM_PSMCT24][PSM_PSMT4HL >> 5] |= 1 << (PSM_PSMT4HL & 0x1f);
+		SharedBitsField[PSM_PSMCT24][PSM_PSMT4HH >> 5] |= 1 << (PSM_PSMT4HH & 0x1f);
+		SharedBitsField[PSM_PSMZ24][PSM_PSMT8H >> 5] |= 1 << (PSM_PSMT8H & 0x1f);
+		SharedBitsField[PSM_PSMZ24][PSM_PSMT4HL >> 5] |= 1 << (PSM_PSMT4HL & 0x1f);
+		SharedBitsField[PSM_PSMZ24][PSM_PSMT4HH >> 5] |= 1 << (PSM_PSMT4HH & 0x1f);
+		SharedBitsField[PSM_PSMT8H][PSM_PSMCT24 >> 5] |= 1 << (PSM_PSMCT24 & 0x1f);
+		SharedBitsField[PSM_PSMT8H][PSM_PSMZ24 >> 5] |= 1 << (PSM_PSMZ24 & 0x1f);
+		SharedBitsField[PSM_PSMT4HL][PSM_PSMCT24 >> 5] |= 1 << (PSM_PSMCT24 & 0x1f);
+		SharedBitsField[PSM_PSMT4HL][PSM_PSMZ24 >> 5] |= 1 << (PSM_PSMZ24 & 0x1f);
+		SharedBitsField[PSM_PSMT4HL][PSM_PSMT4HH >> 5] |= 1 << (PSM_PSMT4HH & 0x1f);
+		SharedBitsField[PSM_PSMT4HH][PSM_PSMCT24 >> 5] |= 1 << (PSM_PSMCT24 & 0x1f);
+		SharedBitsField[PSM_PSMT4HH][PSM_PSMZ24 >> 5] |= 1 << (PSM_PSMZ24 & 0x1f);
+		SharedBitsField[PSM_PSMT4HH][PSM_PSMT4HL >> 5] |= 1 << (PSM_PSMT4HL & 0x1f);
 	}
 
 } s_maps;
@@ -80,21 +85,17 @@ GS_PRIM_CLASS GSUtil::GetPrimClass(uint32 prim)
 
 bool GSUtil::HasSharedBits(uint32 spsm, uint32 dpsm)
 {
-	return s_maps.SharedBitsField[spsm][dpsm];
+	return (s_maps.SharedBitsField[dpsm][spsm >> 5] & (1 << (spsm & 0x1f))) == 0;
 }
 
 bool GSUtil::HasSharedBits(uint32 sbp, uint32 spsm, uint32 dbp, uint32 dpsm)
 {
-	if(sbp != dbp) return false;
-
-	return HasSharedBits(spsm, dpsm);
+	return ((sbp ^ dbp) | (s_maps.SharedBitsField[dpsm][spsm >> 5] & (1 << (spsm & 0x1f)))) == 0;
 }
 
 bool GSUtil::HasCompatibleBits(uint32 spsm, uint32 dpsm)
 {
-	if(spsm == dpsm) return true;
-
-	return s_maps.CompatibleBitsField[spsm][dpsm];
+	return (s_maps.CompatibleBitsField[spsm][dpsm >> 5] & (1 << (dpsm & 0x1f))) != 0;
 }
 
 bool GSUtil::CheckDirectX()

plugins/GSdx/GSVector.h

@@ -1032,7 +1032,7 @@ public:
 	bool allfalse() const
 	{
 		#if _M_SSE >= 0x401
-		return _mm_testz_si128(m, m);
+		return _mm_testz_si128(m, m) != 0;
 		#else
 		return _mm_movemask_epi8(m) == 0;
 		#endif
@@ -2485,15 +2485,15 @@ public:
 
 	GSVector4 clamp(const float scale = 255) const 
 	{
-		return minv(GSVector4(scale));
+		return min(GSVector4(scale));
 	}
 
-	GSVector4 minv(const GSVector4& a) const
+	GSVector4 min(const GSVector4& a) const
 	{
 		return GSVector4(_mm_min_ps(m, a));
 	}
 
-	GSVector4 maxv(const GSVector4& a) const
+	GSVector4 max(const GSVector4& a) const
 	{
 		return GSVector4(_mm_max_ps(m, a));
 	}
@@ -2550,7 +2550,7 @@ public:
 	{
 		#if _M_SSE >= 0x401
 		__m128i a = _mm_castps_si128(m);
-		return _mm_testz_si128(a, a);
+		return _mm_testz_si128(a, a) != 0;
 		#else
 		return _mm_movemask_ps(m) == 0;
 		#endif

plugins/GSdx/GSdx_vs2008.vcproj

@@ -1371,6 +1371,14 @@
 						AssemblerOutput="4"
 					/>
 				</FileConfiguration>
+				<FileConfiguration
+					Name="Debug SSE4|Win32"
+					>
+					<Tool
+						Name="VCCLCompilerTool"
+						AssemblerOutput="4"
+					/>
+				</FileConfiguration>
 				<FileConfiguration
 					Name="Release SSE4|Win32"
 					>

plugins/GSdx/stdafx.h

@@ -190,3 +190,6 @@ typedef signed long long int64;
 	#include <smmintrin.h>
 
 #endif
+
+#undef min
+#undef max